US20080221884A1 - Mobile environment speech processing facility - Google Patents

Abstract

In embodiments of the present invention improved capabilities are described for a mobile environment speech processing facility. The present invention may provide for the entering of text into a software application resident on a mobile communication facility, where recorded speech may be presented by the user using the mobile communications facility's resident capture facility. Transmission of the recording may be provided through a wireless communication facility to a speech recognition facility, and may be accompanied by information related to the software application. Results may be generated utilizing the speech recognition facility that may be independent of structured grammar, and may be based at least in part on the information relating to the software application and the recording. The results may then be transmitted to the mobile communications facility, where they may be loaded into the software application. In embodiments, the user may be allowed to alter the results that are received from the speech recognition facility.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS
• This application claims the benefit of the following provisional applications, each of which is hereby incorporated by reference in its entirety:
• U.S. Provisional App. No. 60/893,600 filed Mar. 7, 2007; and
• U.S. Provisional App. No. 60/976,050 filed Sep. 28, 2007.
BACKGROUND
• 1. Field
• The present invention is related to speech recognition, and specifically to speech recognition in association with a mobile communications facility.
• 2. Description of the Related Art
• Speech recognition, also known as automatic speech recognition, is the process of converting a speech signal to a sequence of words by means of an algorithm implemented as a computer program. Speech recognition applications that have emerged over the last years include voice dialing (e.g., call home), call routing (e.g., I would like to make a collect call), simple data entry (e.g., entering a credit card number), and preparation of structured documents (e.g., a radiology report). Current systems are either not for mobile communication devices or utilize constraints, such as requiring a specified grammar, to provide real-time speech recognition. The current invention provides a facility for unconstrained, mobile, real-time speech recognition.
SUMMARY
• The current invention allows an individual with a mobile communications facility to use speech recognition to enter text, such as into a communications application, such as an SMS message, instant messenger, e-mail, or any other application, such as applications for getting directions, entering query word string into a search engine, commands into a navigation or map program, and a wide range of others.
• In embodiments the present invention may provide for the entering of text into a software application resident on a mobile communication facility, where recorded speech may be presented by the user using the mobile communications facility's resident capture facility. Transmission of the recording may be provided through a wireless communication facility to a speech recognition facility, and may be accompanied by information related to the software application. Results may be generated utilizing the speech recognition facility that may be independent of structured grammar, and may be based at least in part on the information relating to the software application and the recording. The results may then be transmitted to the mobile communications facility, where they may be loaded into the software application. In embodiments, the user may be allowed to alter the results that are received from the speech recognition facility. In addition, the speech recognition facility may be adapted based on usage.
• In embodiments, the information relating to the software application may include at least one of an identity of the application, an identity of a text box within the application, contextual information within the application, an identity of the mobile communication facility, an identity of the user, and the like.
• In embodiments, the step of generating the results may be based at least in part on the information relating to the software application involved in selecting at least one of a plurality of recognition models based on the information relating to the software application and the recording, where the recognition models may include at least one of an acoustic model, a pronunciation, a vocabulary, a language model, and the like, and at least one of a plurality of language models, wherein the at least one of the plurality of language models may be selected based on the information relating to the software application and the recording. In embodiments, the plurality of language models may be run at the same time or in multiple passes in the speech recognition facility. The selection of language models for subsequent passes may be based on the results obtained in previous passes. The output of multiple passes may be combined into a single result by choosing the highest scoring result, the results of multiple passes, and the like, where the merging of results may be at the word, phrase, or the like level.
• In embodiments, adapting the speech recognition facility may be based on usage that includes at least one of adapting an acoustic model, adapting a pronunciation, adapting a vocabulary, adapting a language model, and the like. Adapting the speech recognition facility may include adapting recognition models based on usage data, where the process may be an automated process, the models may make use of the recording, the models may make use of words that are recognized, the models may make use of the information relating to the software application about action taken by the user, the models may be specific to the user or groups of users, the models may be specific to text fields with in the software application or groups of text fields within the software applications, and the like.
• In embodiments, the step of allowing the user to alter the results may include the user editing a text result using at least one of a keypad or a screen-based text correction mechanism, selecting from among a plurality of alternate choices of words contained in the results, selecting from among a plurality of alternate actions related to the results, selecting among a plurality of alternate choices of phrases contained in the results, selecting words or phrases to alter by speaking or typing, positioning a cursor and inserting text at the cursor position by speaking or typing, and the like. In addition, the speech recognition facility may include a plurality of recognition models that may be adapted based on usage, including utilizing results altered by the user, adapting language models based on usage from results altered by the user, and the like.
• These and other systems, methods, objects, features, and advantages of the present invention will be apparent to those skilled in the art from the following detailed description of the preferred embodiment and the drawings. All documents mentioned herein are hereby incorporated in their entirety by reference.
BRIEF DESCRIPTION OF THE FIGURES
• The invention and the following detailed description of certain embodiments thereof may be understood by reference to the following figures:
• FIG. 1 depicts a block diagram of the mobile environment speech processing facility.
• FIG. 2 depicts a block diagram of the automatic speech recognition server infrastructure architecture.
• FIG. 3 depicts a block diagram of the application infrastructure architecture.
• FIG. 4 depicts some of the components of the ASR Client.
• FIG. 5adepicts the process by which multiple language models may be used by the ASR engine.
• FIG. 5bdepicts the process by which multiple language models may be used by the ASR engine for a navigation application embodiment.
• FIG. 5cdepicts the process by which multiple language models may be used by the ASR engine for a messaging application embodiment.
• FIG. 5ddepicts the process by which multiple language models may be used by the ASR engine for a content search application embodiment.
• FIG. 5edepicts the process by which multiple language models may be used by the ASR engine for a search application embodiment.
• FIG. 5fdepicts the process by which multiple language models may be used by the ASR engine for a browser application embodiment.
• FIG. 6 depicts the components of the ASR engine.
• FIG. 7 depicts the layout and initial screen for the user interface.
• FIG. 8 depicts a keypad layout for the user interface.
• FIG. 9 depicts text boxes for the user interface.
• FIG. 10 depicts a first example of text entry for the user interface.
• FIG. 11 depicts a second example of text entry for the user interface.
• FIG. 12 depicts a third example of text entry for the user interface.
• FIG. 13 depicts speech entry for the user interface.
• FIG. 14 depicts speech-result correction for the user interface.
• FIG. 15 depicts a first example of navigating browser screen for the user interface.
• FIG. 16 depicts a second example of navigating browser screen for the user interface.
• FIG. 17 depicts packet types communicated between the client, router, and server at initialization and during a recognition cycle.
• FIG. 18 depicts an example of the contents of a header.
• FIG. 19 depicts the format of a status packet.
DETAILED DESCRIPTION
• The current invention provides an unconstrained, real-time, mobile environmentspeech processing facility100, as shown inFIG. 1, allowing a user with amobile communications facility120 to use speech recognition to enter text into anapplication112, such as a communications application, such as an SMS message, IM message, e-mail, chat, blog, or the like, or any other kind of application, such as a social network application, mapping application, application for obtaining directions, search engine, auction application, application related to music, travel, games, or other digital media, enterprise software applications, word processing, presentation software, and the like. In various embodiments, text obtained through the speech recognition facility described herein may be entered into any application or environment that takes text input.
• In an embodiment of the invention, the user's130mobile communications facility120 may be a mobile phone, programmable through a standard programming language, such as Java, C, or C++. The mobile environmentspeech processing facility100 may include a preloadedmobile communications facility120. Or, theuser130 may download theapplication112 to theirmobile communications facility120. Theapplication112 may be for example anavigation application112, a music player, a music download service, amessaging application112 such as SMS or email, a video player orsearch application112, alocal search application112, amobile search application112, a general internet browser or the like. There may also bemultiple applications112 loaded on themobile communications facility120 at the same time. Theuser130 may activate the mobile environment speech processing facility's100user130 interface software by starting a program included in the mobile environmentspeech processing facility120 or activate it by performing auser130 action, such as pushing a button or a touch screen to collect audio into a domain application. The audio signal may then be recorded and routed over a network to theservers110 of the mobile environmentspeech processing facility100. The text output from theservers110, representing the user's130 spoken words, may then be routed back to the user's130mobile communications facility120 for display. In embodiments, theuser130 may receive feedback from the mobile environmentspeech processing facility100 on the quality of the audio signal, for example, whether the audio signal has the right amplitude; whether the audio signal's amplitude is clipped, such as clipped at the beginning or at the end; whether the signal was too noisy; or the like.
• Theuser130 may correct the returned text with the mobile phone's keypad or touch screen navigation buttons. This process may occur in real-time, creating an environment where a mix of speaking and typing is enabled in combination with other elements on the display. The corrected text may be routed back to theservers110, where theASR Server204Infrastructure102 may use the corrections to help model how auser130 typically speaks, what words they use, how theuser130 tends to use words, in what contexts theuser130 speaks, and the like. Theuser130 may speak or type into text boxes, with keystrokes routed back to theASR server204. The corespeech recognition engine208 may include automated speech recognition (ASR), and may utilize a plurality ofmodels218, such as acoustic models220,pronunciations222,vocabularies224,language models228, and the like, in the analysis and translation ofuser130 inputs.Personal language models228 may be biased for first, last name in an address book, user's130 location, phone number, past usage data, or the like. As a result of this dynamic development ofuser130 speech profiles, theuser130 may be free from constraints on how to speak; there may be no grammatical constraints placed on themobile user130, such as having to say something in a fixed domain. Theuser130 may be able to say anything theuser130 wants into the user's130mobile communications facility120, allowing theuser130 to utilize text messaging, searching, entering an address, or the like, and ‘speaking into’ the text field, rather than having to type everything.
• In addition, the hostedservers110 may be run as an application service provider (ASP). This may allow the benefit of running data frommultiple applications112 andusers130, combining them to make moreeffective recognition models218. This may allow better adaptation to theuser130, to the scenario, and to theapplication112, based on usage.
• FIG. 1 depicts an architectural block diagram for the mobile environmentspeech processing facility100, including amobile communications facility120 and hostedservers110 The ASR client may provide the functionality of speech-enabled text entry to the application. TheASR server infrastructure102 may interface with theASR client118, in the user's130mobile communications facility120, via a data protocol, such as a transmission control protocol (TCP) connection or the like. TheASR server infrastructure102 may also interface with theuser database104. Theuser database104 may also be connected with theregistration108 facility. TheASR server infrastructure102 may make use ofexternal information sources124 to provide information about words, sentences, and phrases that theuser130 is likely to speak. Theapplication112 in the user'smobile communication facility120 may also make use of server-side application infrastructure122, also via a data protocol. The server-side application infrastructure122 may provide content for the applications, such as navigation information, music or videos to download, search facilities for content, local, or general web search, and the like. The server-side application infrastructure122 may also provide general capabilities to the application such as translation of HTML or other web-based markup into a form which is suitable for theapplication112. Within the user's130mobile communications facility120,application code114 may interface with theASR client118 via a resident software interface, such as Java, C, or C++. Theapplication infrastructure122 may also interface with theuser database104, and with other externalapplication information sources128 such as theWorld Wide Web330, or with external application-specific content such as navigation services, music, video, search services, and the like.
• FIG. 2 depicts the architecture for theASR server infrastructure102, containing functional blocks for theASR client118,ASR router202,ASR server204,ASR engine208,recognition models218, usage data212,human transcription210,adaptation process214,external information sources124, anduser130database104. In a typical deployment scenario,multiple ASR servers204 may be connected to anASR router202;many ASR clients118 may be connected tomultiple ASR routers102, and network traffic load balancers may be presented betweenASR clients118 andASR routers202. TheASR client118 may present agraphical user130 interface to theuser130, and establishes a connection with theASR router202. TheASR client118 may pass information to theASR router202, including a unique identifier for the individual phone (client ID) that may be related to auser130 account created during a subscription process, and the type of phone (phone ID). TheASR client118 may collect audio from theuser130. Audio may be compressed into a smaller format. Compression may be a standard compression scheme used for human-human conversation, or a specific compression scheme optimized for speech recognition. Theuser130 may indicate that theuser130 would like to perform recognition. Indication may be made by way of pressing and holding a button for the duration theuser130 is speaking. Indication may be made by way of pressing a button to indicate that speaking will begin, and theASR client118 may collect audio until it determines that theuser130 is done speaking, by determining that there has been no speech within some pre-specified time period. In embodiments, voice activity detection may be entirely automated without the need for an initial key press, such as by voice trained command, by voice command specified on the display of themobile communications facility120, or the like.
• TheASR client118 may pass audio, or compressed audio, to theASR router202. The audio may be sent after all audio is collected or streamed while the audio is still being collected. The audio may include additional information about the state of theASR client118 andapplication112 in which this client is embedded. This additional information, plus the client ID and phone ID, is the client state information. This additional information may include an identifier for the application; an identifier for the particular text field of the application; an identifier for content being viewed in the current application, the URL of the current web page being viewed in a browser for example; or words which are already entered into a current text field. There may be information about what words are before and after the current cursor location, or alternatively, a list of words along with information about the current cursor location. This additional information may also include other information available in theapplication112 ormobile communication facility120 which may be helpful in predicting whatusers130 may speak into theapplication112 such as the current location of the phone, information about content such as music or videos stored on the phone, history of usage of the application, time of day, and the like.
• TheASR client118 may wait for results to come back from theASR router202. Results may be returned as word strings representing the system's hypothesis about the words, which were spoken. The result may include alternate choices of what may have been spoken, such as choices for each word, choices for strings of multiple words, or the like. TheASR client118 may present words to theuser130, that appear at the current cursor position in the text box, or shown to theuser130 as alternate choices by navigating with the keys on themobile communications facility120. TheASR client118 may allow theuser130 to correct text by using a combination of selecting alternate recognition hypotheses, navigating to words, seeing list of alternatives, navigating to desired choice, selecting desired choice; deleting individual characters, using some delete key on the keypad or touch screen; deleting entire words one at a time; inserting new characters by typing on the keypad; inserting new words by speaking; replacing highlighted words by speaking; or the like. The list of alternatives may be alternate words or strings of word, or may make use of application constraints to provide a list of alternate application-oriented items such as songs, videos, search topics or the like. TheASR client118 may also give a user130 a means to indicate that theuser130 would like the application to take some action based on the input text; sending the current state of the input text (accepted text) back to theASR router202 when theuser130 selects the application action based on the input text; logging various information aboutuser130 activity by keeping track ofuser130 actions, such as timing and content of keypad or touch screen actions, or corrections, and periodically sending it to theASR router202; or the like.
• TheASR router202 may provide a connection between theASR client118 and theASR server204. TheASR router202 may wait for connection requests fromASR clients118. Once a connection request is made, theASR router202 may decide whichASR server204 to use for the session from theASR client118. This decision may be based on the current load on eachASR server204; the best predicted load on eachASR server204; client state information; information about the state of eachASR server204, which may includecurrent recognition models218 loaded on theASR engine208 or status of other connections to eachASR server204; information about the best mapping of client state information to server state information; routing data which comes from theASR client118 to theASR server204; or the like. TheASR router202 may also route data, which may come from theASR server204, back to theASR client118.
• TheASR server204 may wait for connection requests from theASR router202. Once a connection request is made, theASR server204 may decide whichrecognition models218 to use given the client state information coming from theASR router202. TheASR server204 may perform any tasks needed to get theASR engine208 ready for recognition requests from theASR router202. This may include pre-loadingrecognition models218 into memory, or doing specific processing needed to get theASR engine208 orrecognition models218 ready to perform recognition given the client state information. When a recognition request comes from theASR router202, theASR server204 may perform recognition on the incoming audio and return the results to theASR router202. This may include decompressing the compressed audio information, sending audio to theASR engine208, getting results back from theASR engine208, optionally applying a process to alter the words based on the text and on the Client State Information (changing “five dollars” to $5 for example), sending resulting recognized text to theASR router202, and the like. The process to alter the words based on the text and on the Client State Information may depend on theapplication112, for example applying address-specific changes (changing “seventeen dunster street to” to “17 dunster st.”) in a location-basedapplication112 such as navigation or local search, applying internet-specific changes (changing “yahoo dot com” to “yahoo.com”) in asearch application112, and the like.
• TheASR server204 may log information to the usage data212 storage. This logged information may include audio coming from theASR router202, client state information, recognized text, accepted text, timing information,user130 actions, and the like. TheASR server204 may also include a mechanism to examine the audio data and decide that thecurrent recognition models218 are not appropriate given the characteristics of the audio data and the client state information. In this case theASR server204 may load new oradditional recognition models218, do specific processing needed to getASR engine208 orrecognition models218 ready to perform recognition given the client state information and characteristics of the audio data, rerun the recognition based on these new models, send back information to theASR router202 based on the acoustic characteristics causing the ASR to send the audio to adifferent ASR server204, and the like.
• TheASR engine208 may utilize a set ofrecognition models218 to process the input audio stream, where there may be a number of parameters controlling the behavior of theASR engine208. These may include parameters controlling internal processing components of theASR engine208, parameters controlling the amount of processing that the processing components will use, parameters controlling normalizations of the input audio stream, parameters controlling normalizations of therecognition models218, and the like. TheASR engine208 may output words representing a hypothesis of what theuser130 said and additional data representing alternate choices for what theuser130 may have said. This may include alternate choices for the entire section of audio; alternate choices for subsections of this audio, where subsections may be phrases (strings of one or more words) or words; scores related to the likelihood that the choice matches words spoken by theuser130; or the like. Additional information supplied by theASR engine208 may relate to the performance of theASR engine208.
• Therecognition models218 may control the behavior of theASR engine208. These models may contain acoustic models220, which may control how theASR engine208 maps the subsections of the audio signal to the likelihood that the audio signal corresponds to each possible sound making up words in the target language. These acoustic models220 may be statistical models, Hidden Markov models, may be trained on transcribed speech coming from previous use of the system (training data), multiple acoustic models with each trained on portions of the training data, models specific tospecific users130 or groups ofusers130, or the like. These acoustic models may also have parameters controlling the detailed behavior of the models. Therecognition models218 may include acoustic mappings, which represent possible acoustic transformation effects, may include multiple acoustic mappings representing different possible acoustic transformations, and these mappings may apply to the feature space of theASR engine208. Therecognition models218 may include representations of thepronunciations222 of words in the target language. Thesepronunciations222 may be manually created by humans, derived through a mechanism which converts spelling of words to likely pronunciations, derived based on spoken samples of the word, and may include multiple possible pronunciations for each word in thevocabulary224, multiple sets of pronunciations for the collection of words in thevocabulary224, and the like. Therecognition models218 may includelanguage models228, which represent the likelihood of various word sequences that may be spoken by theuser130. Theselanguage models228 may be statistical language models, n-gram statistical language models, conditional statistical language models which take into account the client state information, may be created by combining the effects of multiple individual language models, and the like. Therecognition models218 may includemultiple language models228 which are used in a variety of combinations by theASR engine208. Themultiple language models228 may includelanguage models228 meant to represent the likely utterances of aparticular user130 or group ofusers130. Thelanguage models228 may be specific to theapplication112 or type ofapplication112.
• Themultiple language models228 may includelanguage models228 designed to model words, phrases, and sentences used by people speaking destinations for a navigation orlocal search application112 or the like. Thesemultiple language models228 may includelanguage models228 about locations,language models228 about business names,language models228 about business categories,language models228 about points of interest,language models228 about addresses, and the like. Each of these types oflanguage models228 may be general models which provide broad coverage for each of the particular type of ways of entering a destination or may be specific models which are meant to model the particular businesses, business categories, points of interest, or addresses which appear only within a particular geographic region.
• Themultiple language models228 may includelanguage models228 designed to model words, phrases, and sentences used by people speaking intomessaging applications112. Theselanguage models228 may includelanguage models228 specific to addresses, headers, and content fields of amessaging application112. Thesemultiple language models228 may be specific to particular types of messages ormessaging application112 types.
• Themultiple language models228 may includelanguage models228 designed to model words, phrases, and sentences used by people speaking search terms for content such as music, videos, games, and the like. Thesemultiple language models228 may includelanguage models228 representing artist names, song names, movie titles, TV show, popular artists, and the like. Thesemultiple language models228 may be specific to various types of content such as music or video category or may cover multiple categories.
• Themultiple language models228 may includelanguage models228 designed to model words, phrases, and sentences used by people speaking general search terms into a search application. Themultiple language models228 may includelanguage models228 for particular types of search including content search, local search, business search, people search, and the like.
• Themultiple language models228 may includelanguage models228 designed to model words, phrases, and sentences used by people speaking text into a general internet browser. Thesemultiple language models228 may includelanguage models228 for particular types of web pages or text entry fields such as search, form filling, dates, times, and the like.
• Usage data212 may be a stored set of usage data212 from theusers130 of the service that includes stored digitized audio that may be compressed audio; client state information from each audio segment; accepted text from theASR client118; logs ofuser130 behavior, such as key-presses; and the like. Usage data212 may also be the result ofhuman transcription210 of stored audio, such as words that were spoken byuser130, additional information such as noise markers, information about the speaker such as gender or degree of accent, or the like.
• Human transcription210 may be software and processes for a human to listen to audio stored in usage data212, and annotate data with words which were spoken, additional information such as noise markers, truncated words, information about the speaker such as gender or degree of accent, or the like. A transcriber may be presented with hypothesized text from the system or presented with accepted text from the system. Thehuman transcription210 may also include a mechanism to target transcriptions to a particular subset of usage data212. This mechanism may be based on confidence scores of the hypothesized transcriptions from theASR server204.
• Theadaptation process214 may adaptrecognition models218 based on usage data212. Another criterion foradaptation214 may be to reduce the number of errors that theASR engine208 would have made on the usage data212, such as by rerunning the audio through theASR engine208 to see if there is a better match of the recognized words to what theuser130 actually said. Theadaptation214 techniques may attempt to estimate what theuser130 actually said from the annotations of thehuman transcription210, from the accepted text, from other information derived from the usage data212, or the like. Theadaptation214 techniques may also make use ofclient state information514 to producerecognition models218 that are personalized to anindividual user130 or group ofusers130. For a givenuser130 or group ofusers130, thesepersonalized recognition models218 may be created from usage data212 for thatuser130 or group, as well as data fromusers130 outside of the group such as through collaborative-filtering techniques to determine usage patterns from a large group ofusers130. Theadaptation process214 may also make use of application information to adaptrecognition models218 forspecific domain applications112 or text fields withindomain applications112. Theadaptation process214 may make use of information in the usage data212 to adaptmultiple language models228 based on information in the annotations of thehuman transcription210, from the accepted text, from other information derived from the usage data212, or the like. Theadaptation process214 may make use ofexternal information sources124 to adapt therecognition models218. Theseexternal information sources124 may contain recordings of speech, may contain information about the pronunciations of words, may contain examples of words thatusers130 may speak into particular applications, may contain examples of phrases and sentences whichusers130 may speak into particular applications, and may contain structured information about underlying entities or concepts thatusers130 may speak about. Theexternal information sources124 may include databases of location entities including city and state names, geographic area names, zip codes, business names, business categories, points of interest, street names, street number ranges on streets, and other information related to locations and destinations. These databases of location entities may include links between the various entities such as which businesses and streets appear in which geographic locations and the like. Theexternal information124 may include sources of popular entertainment content such as music, videos, games, and the like. Theexternal information124 may include information about popular search terms, recent news headlines, or other sources of information which may help predict what users may speak into aparticular application112. Theexternal information sources124 may be specific to aparticular application112, group ofapplications112,user130, or group ofusers130. Theexternal information sources124 may include pronunciations of words that users may use. Theexternal information124 may include recordings of people speaking a variety of possible words, phrases, or sentences. Theadaptation process214 may include the ability to convert structured information about underlying entities or concepts into words, phrases, or sentences whichusers130 may speak in order to refer to those entities or concepts. Theadaption process214 may include the ability to adapt each of themultiple language models228 based on relevant subsets of theexternal information sources124 and usage data212. Thisadaptation214 oflanguage models228 on subsets ofexternal information source124 and usage data212 may include adapting geographic location-specific language models228 based on location entities and usage data212 from only that geographic location, adapting application-specific language models based on theparticular application112 type,adaptation124 based on related data or usages, or may include adapting124language models228 specific toparticular users130 or groups ofusers130 on usage data212 from just thatuser130 or group ofusers130.
• Theuser database104 may be updated byweb registration108 process, by new information coming from theASR router202, by new information coming from theASR server204, by tracking application usage statistics, or the like. Within theuser database104 there may be two separate databases, the ASR database and theuser database104. The ASR database may contain a plurality of tables, such as asr_servers; asr_routers; asr_am (AM, profile name & min server count); asr_monitor (debugging), and the like. The user130 database104 may also contain a plurality of tables, such as a clients table including client ID, user130 ID, primary user130 ID, phone number, carrier, phone make, phone model, and the like; a users130 table including user130 ID, developer permissions, registration time, last activity time, activity count recent AM ID, recent LM ID, session count, last session timestamp, AM ID (default AM for user130 used from priming), and the like; a user130 preferences table including user130 ID, sort, results, radius, saved searches, recent searches, home address, city, state (for geocoding), last address, city, state (for geocoding), recent locations, city to state map (used to automatically disambiguate one-to-many city/state relationship) and the like; user130 private table including user130 ID, first and last name, email, password, gender, type of user130 (e.g. data collection, developer, VIP, etc), age and the like; user130 parameters table including user130 ID, recognition server URL, proxy server URL, start page URL, logging server URL, logging level, is Logging, isDeveloper, or the like; clients updates table used to send update notices to clients, including client ID, last known version, available version, minimum available version, time last updated, time last reminded, count since update available, count since last reminded, reminders sent, reminder count threshold, reminder time threshold, update URL, update version, update message, and the like; or other similar tables, such as application usage data212 not related to ASR.
• FIG. 3 depicts an example browser-basedapplication infrastructure architecture300 including thebrowser renderer302, the browser proxy604, text-to-speech (TTS)server308,TTS engine310, speech aware mobile portal (SAMP)312, text-box router314,domain applications312,scrapper320,user130database104, and theWorld Wide Web330. Thebrowser renderer302 may be a part of theapplication code114 in the usersmobile communication facility120 and may provide a graphical andspeech user130 interface for theuser130 and display elements on screen-based information coming frombrowser proxy304. Elements may include text elements, image elements, link elements, input elements, format elements, and the like. Thebrowser renderer302 may receive input from theuser130 and send it to thebrowser proxy304. Inputs may include text in a text-box, clicks on a link, clicks on an input element, or the like. Thebrowser renderer302 also may maintain the stack required for “Back” key presses, pages associated with each tab, and cache recently-viewed pages so that no reads from proxy are required to display recent pages (such as “Back”).
• Thebrowser proxy304 may act as an enhanced HTML browser that issues http requests for pages, http requests for links, interprets HTML pages, or the like. Thebrowser proxy304 may convertuser130 interface elements into a form required for thebrowser renderer302. Thebrowser proxy304 may also handle TTS requests from thebrowser renderer302; such as sending text to theTTS server308; receiving audio from theTTS server308 that may be in compressed format; sending audio to thebrowser renderer302 that may also be in compressed format; and the like.
• Other blocks of the browser-basedapplication infrastructure300 may include aTTS server308,TTS engine310,SAMP312,user130 database104 (previously described), theWorld Wide Web330, and the like. TheTTS server308 may accept TTS requests, send requests to theTTS engine310, receive audio from theTTS engine310, send audio to thebrowser proxy304, and the like. TheTTS engine310 may accept TTS requests, generate audio corresponding to words in the text of the request, send audio to theTTS server308, and the like. TheSAMP312 may handle application requests from thebrowser proxy304, behave similar to aweb application330, include a text-box router314, includedomain applications318, include ascrapper320, and the like. The text-box router314 may accept text as input, similar to a search engine's search box, semantically parsing input text using geocoding, key word and phrase detection, pattern matching, and the like. The text-box router314 may also route parse requests accordingly toappropriate domain applications318 or theWorld Wide Web330.Domain applications318 may refer to a number ofdifferent domain applications318 that may interact with content on theWorld Wide Web330 to provide application-specific functionality to the browser proxy. And finally, thescrapper320 may act as a generic interface to obtain information from the World Wide Web330 (e.g., web services, SOAP, RSS, HTML, scrapping, and the like) and formatting it for the small mobile screen.
• FIG. 4 depicts some of the components of theASR Client114. TheASR client114 may include anaudio capture402 component which may wait for signals to begin and end recording, interacts with the built-in audio functionality on themobile communication facility120, interact with theaudio compression408 component to compress the audio signal into a smaller format, and the like. Theaudio capture402 component may establish a data connection over the data network using theserver communications component410 to theASR server infrastructure102 using a protocol such as TCP or HTTP. Theserver communications410 component may then wait for responses from theASR server infrastructure102 indicated words which the user may have spoken. Thecorrection interface404 may display words, phrases, sentences, or the like, to the user,130 indicating what theuser130 may have spoken and may allow theuser130 to correct or change the words using a combination of selecting alternate recognition hypotheses, navigating to words, seeing list of alternatives, navigating to desired choice, selecting desired choice; deleting individual characters, using some delete key on the keypad or touch screen; deleting entire words one at a time; inserting new characters by typing on the keypad; inserting new words by speaking; replacing highlighted words by speaking; or the like.Audio compression408 may compress the audio into a smaller format using audio compression technology built into themobile communication facility120, or by using its own algorithms for audio compression. Theseaudio compression408 algorithms may compress the audio into a format which can be turned back into a speech waveform, or may compress the audio into a format which can be provided to theASR engine208 directly or uncompressed into a format which may be provided to theASR engine208.Server communications410 may use existing data communication functionality built into themobile communication facility120 and may use existing protocols such as TCP, HTTP, and the like.
• FIG. 5adepicts theprocess500aby which multiple language models may be used by the ASR engine. For the recognition of a given utterance, afirst process504 may decide on an initial set oflanguage models228 for the recognition. This decision may be made based on the set of information in theclient state information514, including application ID, user ID, text field ID, current state ofapplication112, or information such as the current location of themobile communication facility120. TheASR engine208 may then run508 using this initial set oflanguage models228 and a set of recognition hypotheses created based on this set oflanguage models228. There may then be adecision process510 to decide if additional recognition passes508 are needed withadditional language models228. Thisdecision510 may be based on theclient state information514, the words in the current set of recognition hypotheses, confidence scores from the most recent recognition pass, and the like. If needed, a new set oflanguage models228 may be determined518 based on theclient state information514 and the contents of the most recent recognition hypotheses and another pass ofrecognition508 made by theASR engine208. Once complete, the recognition results may be combined to form a single set of words and alternates to pass back to theASR client118.
• FIG. 5bdepicts theprocess500bby whichmultiple language models228 may be used by theASR engine208 for anapplication112 which allowsspeech input502 about locations, such as a navigation, local search, ordirectory assistance application112. For the recognition of a given utterance, afirst process522 may decide on an initial set oflanguage models228 for the recognition. This decision may be made based on the set of information in theclient state information524, including application ID, user ID, text field ID, current state ofapplication112, or information such as the current location of themobile communication facility120. This client state information may also include favorites or an address book from theuser130 and may also include usage history for theapplication112. The decision about the initial set oflanguage models228 may be based on likely target cities for thequery522. The initial set oflanguage models228 may includegeneral language models228 about business names, business categories, city and state names, points of interest, street addresses, and other location entities or combinations of these types of location entities. The initial set oflanguage models228 may also includemodels228 for each of the types of location entities specific to one or more geographic regions, where the geographic regions may be based on the phone's current geographic location, usage history for theparticular user130, or other information in thenavigation application112 which may be useful in predicting the likely geographic area theuser130 may want to enter into theapplication112. The initial set oflanguage models228 may also includelanguage models228 specific to theuser130 or group to which theuser130 belongs. TheASR engine208 may then run508 using this initial set oflanguage models228 and a set of recognition hypotheses created based on this set oflanguage models228. There may then be adecision process510 to decide if additional recognition passes508 are needed withadditional language models228. Thisdecision510 may be based on theclient state information524, the words in the current set of recognition hypotheses, confidence scores from the most recent recognition pass, and the like. This decision may include determining the likely geographic area of the utterance and comparing that to the assumed geographic area or set of areas in theinitial language models228. This determining the likely geographic area of the utterance may include looking for words in the hypothesis or set of hypotheses, which may correspond to a geographic region. These words may include names for cities, states, areas and the like or may include a string of words corresponding to a spoken zip code. If needed, a new set oflanguage models228 may be determined528 based on theclient state information524 and the contents of the most recent recognition hypotheses and another pass ofrecognition508 made by theASR engine208. This new set oflanguage models228 may includelanguage models228 specific to a geographic region determined from a hypothesis or set of hypotheses from the previous recognition pass Once complete, the recognition results may be combined512 to form a single set of words and alternates to pass back520 to theASR client118.
• FIG. 5cdepicts theprocess500cby whichmultiple language models228 may be used by theASR engine208 for amessaging application112 such as SMS, email, instant messaging, and the like, forspeech input502. For the recognition of a given utterance, afirst process532 may decide on an initial set oflanguage models228 for the recognition. This decision may be made based on the set of information in theclient state information534, including application ID, user ID, text field ID, or current state ofapplication112. This client state information may include an address book or contact list for the user, contents of the user's messaging inbox and outbox, current state of any text entered so far, and may also include usage history for theapplication112. The decision about the initial set oflanguage models228 may be based on theuser130, theapplication112, the type of message, and the like. The initial set oflanguage models228 may includegeneral language models228 formessaging applications112,language models228 for contact lists and the like. The initial set oflanguage models228 may also includelanguage models228 specific to theuser130 or group to which theuser130 belongs. TheASR engine208 may then run508 using this initial set oflanguage models228 and a set of recognition hypotheses created based on this set oflanguage models228. There may then be adecision process510 to decide if additional recognition passes508 are needed withadditional language models228. Thisdecision510 may be based on theclient state information534, the words in the current set of recognition hypotheses, confidence scores from the most recent recognition pass, and the like. This decision may include determining the type of message entered and comparing that to the assumed type of message or types of messages in theinitial language models228. If needed, a new set oflanguage models228 may be determined538 based on theclient state information534 and the contents of the most recent recognition hypotheses and another pass ofrecognition508 made by theASR engine208. This new set oflanguage models228 may include language models specific to the type of messages determined from a hypothesis or set of hypotheses from the previous recognition pass Once complete, the recognition results may be combined512 to form a single set of words and alternates to pass back520 to theASR client118.
• FIG. 5ddepicts theprocess500dby whichmultiple language models228 may be used by theASR engine208 for acontent search application112 such as music download, music player, video download, video player, game search and download, and the like, forspeech input502. For the recognition of a given utterance, afirst process542 may decide on an initial set oflanguage models228 for the recognition. This decision may be made based on the set of information in theclient state information544, including application ID, user ID, text field ID, or current state ofapplication112. This client state information may include information about the user's content and playlists, either on the client itself or stored in some network-based storage, and may also include usage history for theapplication112. The decision about the initial set oflanguage models228 may be based on theuser130, theapplication112, the type of content, and the like. The initial set oflanguage models228 may includegeneral language models228 for search,language models228 for artists, composers, or performers,language models228 for specific content such as song and album names, movie and TV show names, and the like. The initial set oflanguage models228 may also includelanguage models228 specific to theuser130 or group to which theuser130 belongs. TheASR engine208 may then run508 using this initial set oflanguage models228 and a set of recognition hypotheses created based on this set oflanguage models228. There may then be adecision process510 to decide if additional recognition passes508 are needed withadditional language models228. Thisdecision510 may be based on theclient state information544, the words in the current set of recognition hypotheses, confidence scores from the most recent recognition pass, and the like. This decision may include determining the type of content search and comparing that to the assumed type of content search in theinitial language models228. If needed, a new set oflanguage models228 may be determined548 based on theclient state information544 and the contents of the most recent recognition hypotheses and another pass ofrecognition508 made by theASR engine208. This new set oflanguage models228 may includelanguage models228 specific to the type of content search determined from a hypothesis or set of hypotheses from the previous recognition pass Once complete, the recognition results may be combined512 to form a single set of words and alternates to pass back520 to theASR client118.
• FIG. 5edepicts theprocess500eby whichmultiple language models228 may be used by theASR engine208 for asearch application112 such as general web search, local search, business search, and the like, forspeech input502. For the recognition of a given utterance, afirst process552 may decide on an initial set oflanguage models228 for the recognition. This decision may be made based on the set of information in theclient state information554, including application ID, user ID, text field ID, or current state ofapplication112. This client state information may include information about the phone's location, and may also include usage history for theapplication112. The decision about the initial set oflanguage models228 may be based on theuser130, theapplication112, the type of search, and the like. The initial set oflanguage models228 may includegeneral language models228 for search,language models228 for different types of search such as local search, business search, people search, and the like. The initial set oflanguage models228 may also includelanguage models228 specific to the user or group to which the user belongs. TheASR engine208 may then run508 using this initial set oflanguage models228 and a set of recognition hypotheses created based on this set oflanguage models228. There may then be adecision process510 to decide if additional recognition passes508 are needed withadditional language models228. Thisdecision510 may be based on theclient state information554, the words in the current set of recognition hypotheses, confidence scores from the most recent recognition pass, and the like. This decision may include determining the type of search and comparing that to the assumed type of search in the initial language models. If needed, a new set oflanguage models228 may be determined558 based on theclient state information554 and the contents of the most recent recognition hypotheses and another pass ofrecognition508 made by theASR engine208. This new set oflanguage models228 may includelanguage models228 specific to the type of search determined from a hypothesis or set of hypotheses from the previous recognition pass. Once complete, the recognition results may be combined512 to form a single set of words and alternates to pass back520 to theASR client118.
• FIG. 5fdepicts theprocess500fby whichmultiple language models228 may be used by theASR engine208 for a general browser as a mobile-specific browser or general internet browser forspeech input502. For the recognition of a given utterance, afirst process562 may decide on an initial set oflanguage models228 for the recognition. This decision may be made based on the set of information in theclient state information564, including application ID, user ID, text field ID, or current state ofapplication112. This client state information may include information about the phone's location, the current web page, the current text field within the web page, and may also include usage history for theapplication112. The decision about the initial set oflanguage models228 may be based on theuser130, theapplication112, the type web page, type of text field, and the like. The initial set oflanguage models228 may includegeneral language models228 for search,language models228 for date and time entry,language models228 for digit string entry, and the like. The initial set oflanguage models228 may also includelanguage models228 specific to theuser130 or group to which theuser130 belongs TheASR engine208 may then run508 using this initial set oflanguage models228 and a set of recognition hypotheses created based on this set oflanguage models228. There may then be adecision process510 to decide if additional recognition passes508 are needed withadditional language models228. Thisdecision510 may be based on theclient state information564, the words in the current set of recognition hypotheses, confidence scores from the most recent recognition pass, and the like. This decision may include determining the type of entry and comparing that to the assumed type of entry in theinitial language models228. If needed, a new set oflanguage models228 may be determined568 based on theclient state information564 and the contents of the most recent recognition hypotheses and another pass ofrecognition508 made by theASR engine208. This new set oflanguage models228 may includelanguage models228 specific to the type of entry determined from a hypothesis or set of hypotheses from the previous recognition pass Once complete, the recognition results may be combined512 to form a single set of words and alternates to pass back520 to theASR client118.
• The process to combine recognition output may make use of multiple recognition hypotheses from multiple recognition passes. These multiple hypotheses may be represented as multiple complete sentences or phrases, or may be represented as a directed graph allowing multiple choices for each word. The recognition hypotheses may include scores representing likelihood or confidence of words, phrases, or sentences. The recognition hypotheses may also include timing information about when words and phrases start and stop. The process to combine recognition output may choose entire sentences or phrases from the sets of hypotheses or may construct new sentences or phrases by combining words or fragments of sentences or phrases from multiple hypotheses. The choice of output may depend on the likelihood or confidence scores and may take into account the time boundaries of the words and phrases.
• FIG. 6 shows the components of theASR engine208. The components may includesignal processing602 which may process the input speech either as a speech waveform or as parameters from a speech compression algorithm and create representations which may be used by subsequent processing in theASR engine208. Acoustic scoring604 may use acoustic models220 to determine scores for a variety of speech sounds for portions of the speech input. The acoustic models220 may be statistical models and the scores may be probabilities. Thesearch608 component may make use of the score of speech sounds from theacoustic scoring602 and usingpronunciations222,vocabulary224, andlanguage models228, find the highest scoring words, phrases, or sentences and may also produce alternate choices of words, phrases, or sentences.
• FIG. 7 shows an example of how theuser130 interface layout andinitial screen700 may look on a user's130mobile communications facility120. The layout, from top to bottom, may include a plurality of components, such as a row of navigable tabs, the current page, soft-key labels at the bottom that can be accessed by pressing the left or right soft-keys on the phone, a scroll-bar on the right that shows vertical positioning of the screen on the current page, and the like. The initial screen may contain a text-box with a “Search” button, choices of whichdomain applications318 to launch, a pop-up hint for first-time users130, and the like. The text box may be a shortcut thatusers130 can enter into, or speak into, to jump to adomain application318, such as “Restaurants in Cambridge” or “Send a text message to Joe”. When theuser130 selects the “Search” button, the text content is sent. Application choices may send theuser130 to the appropriate application when selected. The popup hint 1) tells theuser130 to hold the green TALK button to speak, and 2) gives the user130 a suggestion of what to say to try the system out. Both types of hints may go away after several uses.
• Although there are mobile phones with full alphanumeric keyboards, most mass-market devices are restricted to thestandard telephone keypad802, such as shown inFIG. 8. Command keys may include a “TALK”, or green-labeled button, which may be used to make a regular voice-based phone call; an “END” button which is used to terminate a voice-based call or end anapplication112 and go back to the phone's main screen; a five-way control joystick thatusers130 may employ to move up, down, left, and right, or select by pressing on the center button (labeled “MENU/OK” inFIG. 8); two soft-key buttons that may be used to select the labels at the bottom of the screen; a back button which is used to go back to the previous screen in any application; a delete button used to delete entered text that on some phones, such as the one pictured inFIG. 8, the delete and back buttons are collapsed into one; and the like.
• FIG. 9 shows text boxes in a navigate-and-edit mode. A text box is either in navigate mode or editmode900. When in navigatemode902, no cursor or a dim cursor is shown and ‘up/down’, when the text box is highlighted, moves to the next element on the browser screen. For example, moving down would highlight the “search” box. Theuser130 may enter edit mode from navigatemode902 on any of a plurality of actions; including pressing on center joystick; moving left/right in navigate mode; selecting “Edit” soft-key; pressing any of the keys 0-9, which also adds the appropriate letter to the text box at the current cursor position; and the like. When inedit mode904, a cursor may be shown and the left soft-key may be “Clear” rather than “Edit.” The current shift mode may be also shown in the center of the bottom row. Inedit mode904, up and down may navigate within the text box, althoughusers130 may also navigate out of the text box by navigating past the first and last rows. In this example, pressing up would move the cursor to the first row, while pressing down instead would move the cursor out of the text box and highlight the “search” box instead. Theuser130 may hold the navigate buttons down to perform multiple repeated navigations. When the same key is held down for an extended time, four seconds for example, navigation may be sped up by moving more quickly, for instance, times four in speed. As an alternative, navigatemode902 may be removed so that when the text box is highlighted, a cursor may be shown. This may remove the modality, but then requiresusers130 to move up and down through each line of the text box when trying to navigate past the text box.
• Text may be entered in the current cursor position in multi-tap mode, as shown inFIGS. 10,11, and12. As an example, pressing “2” once may be the same as entering “a”, pressing “2” twice may be the same as entering “b”, pressing “2” three times may be the same as entering “c”, and pressing “2” 4 times may be the same as entering “2”. The direction keys may be used to reposition the cursor. Back, or delete on some phones, may be used to delete individual characters. When Back is held down, text may be deleted to the beginning of the previous recognition result, then to the beginning of the text. Capitalized letters may be entered by pressing the “*” key which may put the text into capitalization mode, with the first letter of each new word capitalized. Pressing “*” again puts the text into all-caps mode, with all new entered letters capitalized. Pressing “*” yet again goes back to lower case mode where no new letters may be capitalized. Numbers may be entered either by pressing a key repeatedly to cycle through the letters to the number, or by going into numeric mode. The menu soft-key may contain a “Numbers” option which may put the cursor into numeric mode. Alternatively, numeric mode may be accessible by pressing “*” when cycling capitalization modes. To switch back to alphanumeric mode, theuser130 may again select the Menu soft-key which now contains an “Alpha” option, or by pressing “*”. Symbols may be entered by cycling through the “1” key, which may map to a subset of symbols, or by bringing up the symbol table through the Menu soft-key. The navigation keys may be used to traverse the symbol table and the center OK button used to select a symbol and insert it at the current cursor position.
• FIG. 13 provides examples ofspeech entry1300, and how it is depicted on theuser130 interface. When theuser130 holds the TALK button to begin speaking, a popup may appear informing theuser130 that the recognizer is listening1302. In addition, the phone may either vibrate or play a short beep to cue theuser130 to begin speaking. When theuser130 is finished speaking and releases the TALK button, the popup status may show “Working”1004 with a spinning indicator. Theuser130 may cancel a processing recognition by pressing a button on the keypad or touch screen, such as “Back” or a directional arrow. Finally, when the result is received from theASR server204, the text box may be populated1008.
• When theuser130 presses left or right to navigate through the text box,alternate results1402 for each word may be shown in gray below the cursor for a short time, such as 1.7 seconds. After that period, the gray alternates disappear, and theuser130 may have to move left or right again to get the box. If theuser130 presses down to navigate to the alternates while it is visible, then the current selection in the alternates may be highlighted, and the words that will be replaced in the original sentence may be highlighted in red1404. The image on the bottom left ofFIG. 14 shows a case where two words in the original sentence will be replaced1408. To replace the text with the highlighted alternate, theuser130 may press the center OK key. When the alternate list is shown in red1408 after theuser130 presses down to choose it, the list may become hidden and go back to normal cursor mode if there is no activity after some time, such as 5 seconds. When the alternate list is shown in red, theuser130 may also move out of it by moving up or down past the top or bottom of the list, in which case the normal cursor is shown with no gray alternates box. When the alternate list is shown in red, theuser130 may navigate the text by words by moving left and right. For example, when “nobel” is highlighted1404, moving right would highlight “bookstore” and show its alternate list instead.
• When theuser130 navigates to a new screen, the “Back” key may be used to go back to the previous screen. As shown inFIG. 15, if theuser130 presses “Back” after looking through the search results, the screen on the left is shown1502. When theuser130 navigates to a new page from the home page, a new tab may be automatically inserted to the right of the “home” tab, as shown inFIG. 16. Unless theuser130 is in a text box, tabs can be navigated by pressing left or right keys. Theuser130 may also move to the top of the screen and select the tab itself before moving left or right. When the tab is highlighted, theuser130 may also select the left soft-key to remove the current tab and screen. As an alternative, tabs may show icons instead of names as pictured, tabs may be shown at the bottom of the screen, the initial screen may be pre-populated with tabs, selection of an item from the home page may take theuser130 to an existing tab instead of a new one, and tabs may not be selectable by moving to the top of the screen and tabs may not be removable by theuser130, and the like.
• As shown inFIG. 2, there is communication between theASR client118,ASR router202, andASR server204. These communications may be subject to specific protocols. In these protocols, theASR client118, when prompted byuser130, records audio and sends it to theASR router202. Received results from theASR router202 are displayed for theuser130. Theuser130 may senduser130 entries toASR router202 for any text entry. TheASR router202 sends audio to theappropriate ASR server204, depending on theuser130 profile represented by the client ID and CPU load onASR servers204, then sends the results from theASR server204 back to theASR client118. TheASR router202 re-routes the data if theASR server204 indicates amismatched user130 profile. TheASR router202 sends to theASR server204 anyuser130 text inputs for editing. TheASR server204 receives audio fromASR router202 and performs recognition. Results are returned to theASR router202. TheASR server204 alerts theASR router202 if the user's130 speech no longer matches the user's130 predicteduser130 profile, and theASR router202 handles the appropriate re-route. TheASR server204 also receives user-edit accepted text results from theASR router202.
• FIG. 17 shows an illustration of the packet types that are communicated between theASR client118,ASR router202, andserver204 at initialization and during a recognition cycle. During initialization, a connection is requested, with the connection request going fromASR client118 to theASR router202 and finally to theASR server204. A ready signal is sent back from theASR servers204 to theASR router202 and finally to theASR client118. During the recognition cycle, a waveform is input at theASR client118 and routed to theASR servers204. Results are then sent back out to theASR client118, where theuser130 accepts the returned text, sent back to theASR servers104. A plurality of packet types may be utilized during these exchanges, such as PACKET_WAVEFORM=1, packet is waveform; PACKET_TEXT=2, packet is text; PACKET_END_OF_STREAM=3, end of waveform stream; PACKET_IMAGE=4, packet is image; PACKET_SYNCLIST=5, syncing lists, such as email lists; PACKET_CLIENT_PARAMETERS=6, packet contains parameter updates for client; PACKET_ROUTER_CONTROL=7, packet contains router control information; PACKET_MESSAGE=8, packet contains status, warning or error message; PACKET_IMAGE_REQUEST=9, packet contains request for an image or icon; or the like. In addition, each message may have a header, such as shown inFIG. 18. All multi-byte words are in big-endian format.
• As shown inFIG. 17, initialization may be sent from theASR client118, through theASR router202, to theASR server204. TheASR client118 may open a connection with theASR router202 by sending its Client ID. TheASR router202 in turn looks up the ASR client's118 most recent acoustic model220 (AM) and language model228 (LM) and connects to anappropriate ASR server204. TheASR router202 stores that connection until theASR client118 disconnects or the Model ID changes. The packet format for initialization may have a specific format, such as Packet type=TEXT, Data=ID:<client id string> ClientVersion: <client version string>, Protocol:<protocol id string> NumReconnects: <# attempts client has tried reconnecting to socket>, or the like. The communications path for initialization may be (1) Client sends Client ID toASR router202, (2)ASR router202 forwards to ASR a modified packet: Modified Data=<client's original packet data> SessionCount: <session count string> SpeakerID: <user id sting>\0, and (3) resulting state: ASR is now ready to accept utterance(s) from theASR client118,ASR router202 maintains client's ASR connection.
• As shown inFIG. 17, a ready packet may be sent back to theASR client118 from theASR servers204. The packet format for packet ready may have a specific format, such as Packet type=TEXT, Data=Ready\0, and the communications path may be (1) ASR sends Ready router and (2)ASR router202 forwards Ready packet toASR client118.
• As shown inFIG. 17, a field ID packet containing the name of the application and text field within the application may be sent from theASR client118 to theASR servers204. This packet is sent as soon as theuser130 pushes the TALK button to begin dictating one utterance. TheASR servers204 may use the field ID information to select appropriate recognition models142 for the next speech recognition invocation. TheASR router202 may also use the field ID information to route the current session to adifferent ASR server204. The packet format for the field ID packet may have a specific format, such as Packet type=TEXT; Data=FieldID; <type> <url> <form element name>, for browsing mobile web pages; Data=FieldID: message, for SMS text box; or the like. The connection path may be (1)ASR client118 sends Field ID toASR router202 and (2)ASR router202 forwards to ASR for logging.
• As shown inFIG. 17, a waveform packet may be sent from theASR client118 to theASR servers204. TheASR router202 sequentially streams these waveform packets to theASR server204. If theASR server204 senses a change in the Model ID, it may send the ASR router202 a ROUTER_CONTROL packet containing the new Model ID. In response, theASR router202 may reroute the waveform by selecting an appropriate ASR and flagging the waveform such that thenew ASR server204 will not perform additional computation to generate another Model ID. TheASR router202 may also re-route the packet if the ASR server's204 connection drops or times out. TheASR router202 may keep a cache of the most recent utterance, session information such as the client ID and the phone ID, and corresponding FieldID, in case this happens. The packet format for the waveform packet may have a specific format, such as Packet type=WAVEFORM; Data=audio; with the lower 16 bits of flags set to current Utterance ID of the client. The very first part of WAVEFORM packet may determine the waveform type, currently only supporting AMR or QCELP, where “#!AMR\n” corresponds to AMR and “RIFF” corresponds to QCELP. The connection path may be (1)ASR client118 sends initial audio packet (referred to as the BOS, or beginning of stream) to theASR router202, (2)ASR router202 continues streaming packets (regardless of their type) to the current ASR until one of the following events occur: (a)ASR router202 receives packet type END_OF_STREAM, signaling that this is the last packet for the waveform, (b) ASR disconnects or times out, in whichcase ASR router202 finds new ASR, repeats above handshake, sends waveform cache, and continues streaming waveform from client to ASR until receives END_OF_STREAM, (c) ASR sends ROUTER_CONTROL toASR router202 instructing theASR router202 that the Model ID for that utterance has changed, in which case theASR router202 behaves as in ‘b’, (d) ASRclient118 disconnects or times out, in which case the session is closed, or the like. If the recognizer times out or disconnects after the waveform is sent then theASR router202 may connect to a new ASR.
• As shown inFIG. 17, a request model switch for utterance packet may be sent from theASR server204 to theASR router202. This packet may be sent when theASR server204 needs to flag that itsuser130 profile does not match that of the utterance, i.e. Model ID for the utterances has changed. The packet format for the request model switch for utterance packet may have a specific format, such as Packet type=ROUTER_CONTROL; Data=SwitchModelID: AM=<integer> LM=<integer> SessionID=<integer> UttID=<integer>. The communication may be (1)ASR server204 sends control packet toASR router202 after receiving the first waveform packet, and before sending the results packet, and (2)ASR router202 then finds an ASR which best matches the new Model ID, flags the waveform data such that thenew ASR server204 will not send another SwitchModelID packet, and resends the waveform. In addition, several assumptions may be made for this packet, such as theASR server204 may continue to read the waveform packet on the connection, send a Alternate String or SwitchModelID for every utterance with BOS, and theASR router202 may receive a switch model id packet, it sets the flags value of the waveform packets to <flag value> & 0x8000 to notify ASR that this utterance's Model ID does not need to be checked.
• As shown inFIG. 17, a done packet may be sent from theASR server204 to theASR router202. This packet may be sent when theASR server204 has received the last audio packet, such as type END_OF_STREAM. The packet format for the done packet may have a specific format, such as Packet type=TEXT; with the lower 16 bits of flags set to Utterance ID and Data=Done\0. The communications path may be (1) ASR sends done toASR router202 and (2)ASR router202 forwards toASR client118, assuming theASR client118 only receives one done packet per utterance.
• As shown inFIG. 17, an utterance results packet may be sent from theASR server204 to theASR client118. This packet may be sent when theASR server204 gets a result from theASR engine208. The packet format for the utterance results packet may have a specific format, such as Packet type=TEXT, with the lower 16 bits of flags set to Utterance ID and Data=ALTERNATES: <utterance result string>. The communications path may be (1) ASR sends results toASR router202 and (2)ASR router202 forwards toASR client118. TheASR client118 may ignore the results if the Utterance ID does not match that of the current recognition
• As shown inFIG. 17, an accepted text packet may be sent from theASR client118 to theASR server204. This packet may be sent when theuser130 submits the results of a text box, or when the text box looses focus, as in the API, so that the recognizer can adapt to corrected input as well as fully-texted input. The packet format for the accepted text packet may have a specific format, such as Packet type=TEXT, with the lower 16 bits of flags set to most recent Utterance ID, with Data=Accepted_Text: <accepted utterance string>. The communications path may be (1)ASR client118 sends the text submitted by theuser130 toASR router202 and (2)ASR router202 forwards toASR server204 which recognized results, where <accepted utterance string> contains the text string entered into the text box. In embodiments, other logging information, such as timing information anduser130 editing keystroke information may also be transferred.
• Router control packets may be sent between theASR client118,ASR router202, andASR servers204, to help control theASR router202 during runtime. One of a plurality of router control packets may be a get router status packet. The packet format for the get router status packet may have a specific format, such as Packet type=ROUTER_CONTROL, with Data=GetRouterStatus\0. The communication path may be (1) entity sends this packet to theASR router202 and (2)ASR router202 may respond with a status packet with a specific format, such as theformat1900 shown inFIG. 19.
• Another of a plurality of router control packets may be a busy out ASR server packet. The packet format for the busy out ASR server packet may have a specific format, such as Packet type=ROUTER_CONTROL, with Data=BusyOutASRServer: <ASR Server ID>\0. Upon receiving the busy out ASR server packet, theASR router202 may continue to finish up the existing sessions between theASR router202 and theASR server204 identified by the <ASR Server ID>, and theASR router202 may not start a new session with the saidASR server204. Once all existing sessions are finished, theASR router202 may remove the saidASR server204 from its ActiveServer array. The communication path may be (1) entity sends this packet to theASR router202 and (2)ASR router202 responds with ACK packet with the following format: Packet type=TEXT, and Data=ACK\0.
• Another of a plurality of router control packets may be an immediately remove ASR server packet. The packet format for the immediately remove ASR server packet may have a specific format, such as Packet type=ROUTER_CONTROL, with Data=RemoveASRServer: <ASR Server ID>\0. Upon receiving the immediately remove ASR server packet, theASR router202 may immediately disconnect all current sessions between theASR router202 and theASR server204 identified by the <ASR Server ID>, and theASR router202 may also immediately remove the saidASR server204 from its Active Server array. The communication path may be (1) entity sends this packet to theASR router202 and (2)ASR router202 responds with ACK packet with the following format: Packet type=TEXT, and Data=ACK\0.
• Another of a plurality of router control packets may be an add of anASR server204 to the router packet. When anASR server204 is initially started, it may send the router(s) this packet. TheASR router202 in turn may add thisASR server204 to its Active Server array after establishing thisASR server204 is indeed functional. The packet format for the add anASR server204 to theASR router202 may have a specific format, such as Packet type=ROUTER_CONTROL, with Data=AddASRServer: ID=<server id> IP=<server ip address> PORT=<server port> AM=<server AM integer> LM=<server LM integer> NAME=<server name string> PROTOCOL=<server protocol float>. The communication path may be (1) entity sends this packet to theASR router202 and (2)ASR router202 responds with ACK packet with the following format: Packet type=TEXT, and Data=ACK\0.
• Another of a plurality of router control packets may be an alter router logging format packet. This function may cause theASR router202 to read a logging.properties file, and update its logging format during runtime. This may be useful for debugging purposes. The location of the logging.properties file may be specified when theASR router202 is started. The packet format for the alter router logging format may have a specific format, such as Packet type=ROUTER_CONTROL, with Data=ReadLogConfigurationFile. The communications path may be (1) entity sends this packet to theASR router202 and (2)ASR router202 responds with ACK packet with the following format: Packet type=TEXT, and Data=ACK\0.
• Another of a plurality of router control packets may be a get ASR server status packet. TheASR server204 may self report the status of thecurrent ASR server204 with this packet. The packet format for theget ASR server204 status may have a specific format, such as Packet type=ROUTER_CONTROL, with data=RequestStatus\0. The communications path may be (1) entity sends this packet to theASRServer204 and (2)ASR Server204 responds with a status packet with the following format: Packet type=TEXT; Data=ASRServerStatus: Status=<1 for ok or 0 for error> AM=<AM id> LM=<LM id> NumSessions=<number of active sessions> NumUtts=<number of queued utterances> TimeSinceLastRec=<seconds since last recognizer activity>\n Session: client=<client id> speaker=<speaker id> sessioncount=<sessioncount>\n <other Session: line if other sessions exist>\n \0. This router control packet may be used by theASR router202 when establishing whether or not anASR server204 is indeed functional.
• There may be a plurality of message packets associated with communications between theASR client118,ASR router202, andASR servers204, such as error, warning, and status. The error message packet may be associated with an irrecoverable error, the warning message packet may be associated with a recoverable error, and a status message packet may be informational. All three types of messages may contain strings of the format: “<messageType><message>message</message><cause>cause</cause><code>code</code></messageType>”. “messageType” is one of either “status,” “warning,” or “error.” “message” is intended to be displayed to the user, “cause” is intended for debugging, and “code” is intended to trigger additional actions by the receiver of the message.
• The error packet may be sent when a non-recoverable error occurs and is detected. After an error packet has been sent, the connection may be terminated in 5 seconds by the originator if not already closed by the receiver. The packet format for error may have a specific format, such as Packet type=MESSAGE; and Data=“<error><message>error message</message><cause>error cause</cause><code>error code</code></error>”. The communication path from ASR client118 (the originator) to ASR server204 (the receiver) may be (1)ASR client118 sends error packet toASR server204, (2)ASR server204 should close connection immediately and handle error, and (3)ASR client118 will close connection in 5 seconds if connection is still live. There are a number of potential causes for the transmission of an error packet, such as the ASR has received beginning of stream (BOS), but has not received end of stream (EOS) or any waveform packets for 20 seconds; a client has received corrupted data; theASR server204 has received corrupted data; and the like. Examples of corrupted data may be invalid packet type, checksum mismatch, packet length greater than maximum packet size, and the like.
• The warning packet may be sent when a recoverable error occurs and is detected. After a warning packet has been sent, the current request being handled may be halted. The packet format for warning may have a specific format, such as Packet type=MESSAGE; Data=“<warning><message>warning message</message><cause>warning cause</cause><code>warning code</code></warning>”. The communications path fromASR client118 toASR server204 may be (1)ASR client118 sends warning packet toASR server204 and (2)ASR server204 should immediately handle warning. The communications path fromASR server204 toASR client118 may be (1)ASR server204 sends error packet toASR client118 and (2)ASR client118 should immediately handle warning. There are a number of potential causes for the transmission of a warning packet; such as there are noavailable ASR servers204 to handle the request ModelID because theASR servers204 are busy.
• The status packets may be informational. They may be sent asynchronously and do not disturb any processing requests. The packet format for status may have a specific format, such as Packet type=MESSAGE; Data=“<status><message>status message</message><cause>status cause</cause><code>status code</code></status>”. The communications path fromASR client118 toASR server204 may be (1)ASR client118 sends status packet toASR server204 and (2)ASR server204 should handle status. The communication path fromASR server204 toASR client118 may be (1)ASR server204 sends status packet toASR client118 and (2)ASR client118 should handle status. There are a number of potential causes for the transmission of a status packet, such as anASR server204 detects a model ID change for a waveform, server timeout, server error, and the like.
• The elements depicted in flow charts and block diagrams throughout the figures imply logical boundaries between the elements. However, according to software or hardware engineering practices, the depicted elements and the functions thereof may be implemented as parts of a monolithic software structure, as standalone software modules, or as modules that employ external routines, code, services, and so forth, or any combination of these, and all such implementations are within the scope of the present disclosure. Thus, while the foregoing drawings and description set forth functional aspects of the disclosed systems, no particular arrangement of software for implementing these functional aspects should be inferred from these descriptions unless explicitly stated or otherwise clear from the context.
• Similarly, it will be appreciated that the various steps identified and described above may be varied, and that the order of steps may be adapted to particular applications of the techniques disclosed herein. All such variations and modifications are intended to fall within the scope of this disclosure. As such, the depiction and/or description of an order for various steps should not be understood to require a particular order of execution for those steps, unless required by a particular application, or explicitly stated or otherwise clear from the context.
• The methods or processes described above, and steps thereof, may be realized in hardware, software, or any combination of these suitable for a particular application. The hardware may include a general-purpose computer and/or dedicated computing device. The processes may be realized in one or more microprocessors, microcontrollers, embedded microcontrollers, programmable digital signal processors or other programmable device, along with internal and/or external memory. The processes may also, or instead, be embodied in an application specific integrated circuit, a programmable gate array, programmable array logic, or any other device or combination of devices that may be configured to process electronic signals. It will further be appreciated that one or more of the processes may be realized as computer executable code created using a structured programming language such as C, an object oriented programming language such as C++, or any other high-level or low-level programming language (including assembly languages, hardware description languages, and database programming languages and technologies) that may be stored, compiled or interpreted to run on one of the above devices, as well as heterogeneous combinations of processors, processor architectures, or combinations of different hardware and software.
• Thus, in one aspect, each method described above and combinations thereof may be embodied in computer executable code that, when executing on one or more computing devices, performs the steps thereof. In another aspect, the methods may be embodied in systems that perform the steps thereof, and may be distributed across devices in a number of ways, or all of the functionality may be integrated into a dedicated, standalone device or other hardware. In another aspect, means for performing the steps associated with the processes described above may include any of the hardware and/or software described above. All such permutations and combinations are intended to fall within the scope of the present disclosure.
• While the invention has been disclosed in connection with the preferred embodiments shown and described in detail, various modifications and improvements thereon will become readily apparent to those skilled in the art. Accordingly, the spirit and scope of the present invention is not to be limited by the foregoing examples, but is to be understood in the broadest sense allowable by law.
• All documents referenced herein are hereby incorporated by reference.

Claims (25)

1. A method of entering text into a software application resident on a mobile communication facility comprising:

recording speech presented by a user using a mobile communication facility resident capture facility;

transmitting the recording through a wireless communication facility to a speech recognition facility;

transmitting information relating to the software application to the speech recognition facility;

generating results utilizing the speech recognition facility independent of a structured grammar based at least in part on the information relating to the software application and the recording;

transmitting the results to the mobile communications facility;

allowing the user to alter the results; and

loading the results into the software application.

2. The method ofclaim 1, wherein the step of allowing the user to alter the results includes allowing the user to edit a text result using at least one of a keypad or a screen-based text correction mechanism on the mobile communication facility.

3. The method ofclaim 1, wherein the step of allowing the user to alter the results includes allowing the user to select from among a plurality of alternate choices of words contained in the results from the speech recognition facility.

4. The method ofclaim 1, wherein the step of allowing the user to alter the results includes allowing the user to select from among a plurality of alternate actions related to the results from the speech recognition facility.

5. The method ofclaim 1, wherein the step of allowing the user to alter the results includes allowing the user to select among a plurality of alternate choices of phrases contained in the results from the speech recognition facility.

6. The method ofclaim 1, wherein the step of allowing the user to alter the results includes allowing the user to select words or phrases to alter by speaking or typing.

7. The method ofclaim 1, wherein the step of allowing the user to alter the results includes allowing the user to position a cursor and inserting text at the cursor position by speaking or typing.

8. The method ofclaim 1, wherein the speech recognition facility includes a plurality of recognition models that are adapted based on usage.

9. The method ofclaim 8, wherein the step of being adapted based on usage includes utilizing results altered by the user.

10. The method ofclaim 8, wherein the step of being adapted based on usage includes adapting language models based at least in part on usage from results altered by the user.

11. A system for entering text into a software application resident on a mobile communication facility comprising:

recording speech presented by a user using a mobile communication facility resident capture facility;

transmitting the recording through a wireless communication facility to a speech recognition facility;

transmitting information relating to the software application to the speech recognition facility;

generating results utilizing the speech recognition facility independent of a structured grammar based at least in part on the information relating to the software application and the recording;

transmitting the results to the mobile communications facility;

allowing the user to alter the results; and

loading the results into the software application.

12. The system ofclaim 11, wherein the user is allowed to alter the results, including the user editing a text result using at least one of a keypad or a screen-based text correction mechanism on the mobile communication facility.

13. The system ofclaim 11, wherein the user is allowed to alter the results, including the user selecting from among a plurality of alternate choices of words contained in the results from the speech recognition facility.

14. The system ofclaim 11, wherein the user is allowed to alter the results, including the user selecting from among a plurality of alternate actions related to the results from the speech recognition facility.

15. The system ofclaim 11, wherein the user is allowed to alter the results, including the user selecting among a plurality of alternate choices of phrases contained in the results from the speech recognition facility.

16. The system ofclaim 11, wherein the user is allowed to alter the results, including the user selecting words or phrases to alter by speaking or typing.

17. The system ofclaim 11, wherein the user is allowed to alter the results, including the user positioning a cursor and inserting text at the cursor position by speaking or typing.

18. The system ofclaim 11, wherein the speech recognition facility includes a plurality of recognition models that are adapted based on usage.

19. The system ofclaim 18, wherein the step of being adapted based on usage includes utilizing results altered by the user.

20. The system ofclaim 18, wherein the step of being adapted based on usage includes adapting language models based at least in part on usage from results altered by the user.

21. A system, comprising:

A mobile communication device capable of recording speech and running a resident software module;

a speech recognition facility remote from a mobile communication facility;

a communications facility for transmitting recorded speech and information relating to the software module to the speech recognition facility;

wherein the speech recognition facility generates results by processing the recorded speech independent of a structured grammar and based at least in part on the information relating to the software application.

22. The system ofclaim 21, wherein the communication facility transmits results to the mobile communications device.

23. The system ofclaim 21, wherein results are loaded into the software application on the mobile communications device.

24. The system ofclaim 21, wherein generating the results involves selecting a language model based on the information relating to the software application.

25. A method of entering text into a software application resident on a mobile communication facility comprising:

recording speech presented by a user using a mobile communication facility resident capture facility;

transmitting the recording through a wireless communication facility to a speech recognition facility;

identifying the software application to the speech recognition facility; and

generating results using the speech recognition facility, wherein the speech recognition facility is independent of a structured grammar and wherein the output of the speech recognition facility depends on the identity of the software application.

Images