US20040148171A1 - Method and apparatus for speech synthesis without prosody modification - Google Patents

Abstract

A speech synthesizer is provided that concatenates stored samples of speech units without modifying the prosody of the samples. The present invention is able to achieve a high level of naturalness in synthesized speech with a carefully designed training speech corpus by storing samples based on the prosodic and phonetic context in which they occur. In particular, some embodiments of the present invention limit the training text to those sentences that will produce the most frequent sets of prosodic contexts for each speech unit. Further embodiments of the present invention also provide a multi-tier selection mechanism for selecting a set of samples that will produce the most natural sounding speech.

Description

REFERENCE TO RELATED APPLICATION
• The present application is a continuation-in-part of and claims priority of U.S. patent application Ser. No. 09/850,527, filed May 7, 2001, entitled “Method and Apparatus for Speech Synthesis without Prosody Modification”, which is based on and claims the benefit of U.S. Provisional application having serial No. 60/251,167, filed on Dec. 4, 2000 and entitled “PROSODIC WORD SEGMENTATION AND MULTI-TIER NON-UNIFORM UNIT SELECTION”.[0001]
BACKGROUND OF THE INVENTION
• The present invention relates to speech synthesis. In particular, the present invention relates to prosody in speech synthesis.[0002]
• Text-to-speech technology allows computerized systems to communicate with users through synthesized speech. The quality of these systems is typically measured by how natural or human-like the synthesized speech sounds.[0003]
• Very natural sounding speech can be produced by simply replaying a recording of an entire sentence or paragraph of speech. However, the complexity of human languages and the limitations of computer storage may make it impossible to store every conceivable sentence that may occur in a text. Because of this, the art has adopted a concatenative approach to speech synthesis that can be used to generate speech from any text. This concatenative approach combines stored speech samples representing small speech units such as phonemes, diphones, triphones, or syllables to form a larger speech signal.[0004]
• One problem with such concatenative systems is that a stored speech sample has a pitch and duration that is set by the context in which the sample was spoken. For example, in the sentence “Joe went to the store” the speech units associated with the word “store” have a lower pitch than in the question “Joe went to the store?” Because of this, if stored samples are simply retrieved without reference to their pitch or duration, some of the samples will have the wrong pitch and/or duration for the sentence resulting in unnatural sounding speech.[0005]
• One technique for overcoming this is to identify the proper pitch and duration for each sample. Based on this prosody information, a particular sample may be selected and/or modified to match the target pitch and duration.[0006]
• Identifying the proper pitch and duration is known as prosody prediction. Typically, it involves generating a model that describes the most likely pitch and duration for each speech unit given some text. The result of this prediction is a set of numerical targets for the pitch and duration of each speech segment.[0007]
• These targets can then be used to select and/or modify a stored speech segment. For example, the targets can be used to first select the speech segment that has the closest pitch and duration to the target pitch and duration. This segment can then be used directly or can be further modified to better match the target values.[0008]
• For example, one prior art technique for modifying the prosody of speech segments is the so-called Time-Domain Pitch-Synchronous Overlap-and-Add (TD-PSOLA) technique, which is described in “Pitch-Synchronous Waveform Processing Techniques for Text-to-Speech Synthesis using Diphones”, E. Moulines and F. Charpentier, Speech Communication, vol. 9, no. 5, pp. 453-467, 1990. Using this technique, the prior art increases the pitch of a speech segment by identifying a section of the speech segment responsible for the pitch. This section is a complex waveform that is a sum of sinusoids at multiples of a fundamental frequency F[0009]₀. The pitch period is defined by the distance between two pitch peaks in the waveform.
• To increase the pitch, the prior art copies a segment of the complex waveform that is as long as the pitch period. This copied segment is then shifted by some portion of the pitch period and reinserted into the waveform. For example, to double the pitch, the copied segment would be shifted by one-half the pitch period, thereby inserting a new peak half-way between two existing peaks and cutting the pitch period in half.[0010]
• To lengthen a speech segment, the prior art copies a section of the speech segment and inserts the copy into the complex waveform. In other words, the entire portion of the speech segment after the copied segment is time-shifted by the length of the copied section so that the duration of the speech unit increases.[0011]
• Unfortunately, these techniques for modifying the prosody of a speech unit have not produced completely satisfactory results. In particular, these modification techniques tend to produce mechanical or “buzzy” sounding speech.[0012]
• Thus, it would be desirable to be able to select a stored unit that provides good prosody without modification. However, because of memory limitations, samples cannot be stored for all of the possible prosodic contexts in which a speech unit may be used. Instead, a limited set of samples must be selected for storage. Because of this, the performance of a system that uses stored samples without prosody modification is dependent on what samples are stored.[0013]
• Thus, there is an ongoing need for improving the selection of these stored samples in systems that do not modify the prosody of the stored samples. There is also an ongoing need to reduce the computational complexity associated with identifying the proper prosody for the speech units.[0014]
SUMMARY OF THE INVENTION
• A speech synthesizer is provided that concatenates stored samples of speech units without modifying the prosody of the samples. The present invention is able to achieve a high level of naturalness in synthesized speech with a carefully designed speech corpus by storing samples based on the prosodic and phonetic context in which they occur. In particular, some embodiments of the present invention limit the training text to those sentences that will produce the most frequent sets of prosodic contexts for each speech unit. Further embodiments of the present invention also provide a multi-tier selection mechanism for selecting a set of samples that will produce the most natural sounding speech.[0015]
• Under those embodiments that limit the training text, only a limited set of the sentences in a very large corpus are selected and read by a human into a training speech corpus from which samples of units are selected to produce natural sounding speech. To identify which sentences are to be read, embodiments of the present invention determine a frequency of occurrence for each context vector associated with a speech unit. Context vectors with a frequency of occurrence that is larger than a certain threshold are identified as necessary context vectors. Sentences that include the most necessary context vectors are selected for recording until all of the necessary context vectors have been included in the selected sub-set of sentences.[0016]
• In embodiments that use a multi-tier selection method, a set of candidate speech segments is identified for each speech unit by comparing the input context vector to the context vectors associated with the speech segments. A path through the candidate speech segments is then selected based on differences between the input context vectors and the stored context vectors as well as some cost function that indicates the prosodic smoothness of the resulting concatenated speech signal. Under one embodiment, the cost function gives preference to selecting a series of speech segments that appeared next to each other in the training corpus. In further embodiments, the cost function cost function comprises linear, and/or, one or more higher order coordinates being combinations of at least two factors from a set of factors. The set of factors include: an indication of a position of a speech unit in a phrase; an indication of a position of a speech unit in a word; an indication of a category for a phoneme preceding a speech unit; an indication of a category for a phoneme following a speech unit; an indication of a category for tonal identity of the current speech unit; an indication of a category for tonal identity of a preceding speech unit; an indication of a category for tonal identity of a following speech unit; an indication of a level of stress of a speech unit; an indication of a coupling degree of pitch, duration and/or energy with a neighboring unit; and an indication of a degree of spectral mismatch with a neighboring speech unit. A technique for obtaining an optimized cost function, which can be used in a speech synthesizer as the criterion for the unit selection, is also provided below.[0017]
BRIEF DESCRIPTION OF THE DRAWINGS
• FIG. 1 is a block diagram of a general computing environment in which the present invention may be practiced.[0018]
• FIG. 2 is a block diagram of a mobile device in which the present invention may be practiced.[0019]
• FIG. 3 is a block diagram of a speech synthesis system.[0020]
• FIG. 4 is a block diagram of a system for selecting a training text subset from a very large training corpus.[0021]
• FIG. 5 is a flow diagram for constructing a decision tree under one embodiment of the present invention.[0022]
• FIG. 6 is a block diagram of a multi-tier selection system for selecting speech segments under embodiments of the present invention.[0023]
• FIG. 7 is a flow diagram of a multi-tier selection system for selecting speech segments under embodiments of the present invention.[0024]
• FIG. 8 is a flow diagram for estimating mean opinion score from a context vector or an objective measure.[0025]
• FIG. 9 is a plot of a relationship between mean opinion score and the objective measure.[0026]
• FIG. 10 is a flow diagram illustrating a method for optimizing the objective measure.[0027]
• FIG. 11 is a flow diagram illustrating an exemplary method of optimizing the objective measure.[0028]
DETAILED DESCRIPTION OF ILLUSTRATIVE EMBODIMENTS
• FIG. 1 illustrates an example of a suitable[0029]computing system environment100 on which the invention may be implemented. Thecomputing system environment100 is only one example of a suitable computing environment and is not intended to suggest any limitation as to the scope of use or functionality of the invention. Neither should thecomputing environment100 be interpreted as having any dependency or requirement relating to any one or combination of components illustrated in theexemplary operating environment100.
• The invention is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well known computing systems, environments, and/or configurations that may be suitable for use with the invention include, but are not limited to, personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and the like.[0030]
• The invention may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices. Tasks performed by the programs and modules are described below and with the aid of figures. Those skilled in the art can implement the description and figures as processor executable instructions, which can be written on any form of a computer readable media.[0031]
• With reference to FIG. 1, an exemplary system for implementing the invention includes a general-purpose computing device in the form of a[0032]computer110. Components ofcomputer110 may include, but are not limited to, aprocessing unit120, asystem memory130, and asystem bus121 that couples various system components including the system memory to theprocessing unit120. Thesystem bus121 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. By way of example, and not limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus also known as Mezzanine bus.
• [0033]Computer110 typically includes a variety of computer readable media. Computer readable media can be any available media that can be accessed bycomputer110 and includes both volatile and nonvolatile media, removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed bycomputer100.
• Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, FR, infrared and other wireless media. Combinations of any of the above should also be included within the scope of computer readable media.[0034]
• The[0035]system memory130 includes computer storage media in the form of volatile and/or nonvolatile memory such as read only memory (ROM)131 and random access memory (RAM)132. A basic input/output system133 (BIOS), containing the basic routines that help to transfer information between elements withincomputer110, such as during start-up, is typically stored inROM131.RAM132 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by processingunit120. By way of example, and not limitation, FIG. 1 illustratesoperating system134,application programs135,other program modules136, andprogram data137.
• The[0036]computer110 may also include other removable/non-removable volatile/nonvolatile computer storage media. By way of example only, FIG. 1 illustrates ahard disk drive141 that reads from or writes to non-removable, nonvolatile magnetic media, amagnetic disk drive151 that reads from or writes to a removable, nonvolatile magnetic disk152, and anoptical disk drive155 that reads from or writes to a removable, nonvolatileoptical disk156 such as a CD ROM or other optical media. Other removable/non-removable, volatile/nonvolatile computer storage media that can be used in the exemplary operating environment include, but are not limited to, magnetic tape cassettes, flash memory cards, digital versatile disks, digital video tape, solid state RAM, solid state ROM, and the like. Thehard disk drive141 is typically connected to thesystem bus121 through a non-removable memory interface such asinterface140, andmagnetic disk drive151 andoptical disk drive155 are typically connected to thesystem bus121 by a removable memory interface, such asinterface150.
• The drives and their associated computer storage media discussed above and illustrated in FIG. 1, provide storage of computer readable instructions, data structures, program modules and other data for the[0037]computer110. In FIG. 1, for example,hard disk drive141 is illustrated as storingoperating system144,application programs145,other program modules146, andprogram data147. Note that these components can either be the same as or different fromoperating system134,application programs135,other program modules136, andprogram data137.Operating system144,application programs145,other program modules146, andprogram data147 are given different numbers here to illustrate that, at a minimum, they are different copies.
• A user may enter commands and information into the[0038]computer110 through input devices such as akeyboard162, amicrophone163, and apointing device161, such as a mouse, trackball or touch pad. Other input devices (not shown) may include a joystick, game pad, satellite dish, scanner, or the like. These and other input devices are often connected to theprocessing unit120 through auser input interface160 that is coupled to the system bus, but may be connected by other interface and bus structures, such as a parallel port, game port or a universal serial bus (USB). Amonitor191 or other type of display device is also connected to thesystem bus121 via an interface, such as avideo interface190. In addition to the monitor, computers may also include other peripheral output devices such asspeakers197 andprinter196, which may be connected through an outputperipheral interface190.
• The[0039]computer110 may operate in a networked environment using logical connections to one or more remote computers, such as aremote computer180. Theremote computer180 may be a personal computer, a hand-held device, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above relative to thecomputer110. The logical connections depicted in FIG. 1 include a local area network (LAN)171 and a wide area network (WAN)173, but may also include other networks. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.
• When used in a LAN networking environment, the[0040]computer110 is connected to theLAN171 through a network interface oradapter170. When used in a WAN networking environment, thecomputer110 typically includes amodem172 or other means for establishing communications over theWAN173, such as the Internet. Themodem172, which may be internal or external, may be connected to thesystem bus121 via theuser input interface160, or other appropriate mechanism. In a networked environment, program modules depicted relative to thecomputer110, or portions thereof, may be stored in the remote memory storage device. By way of example, and not limitation, FIG. 1 illustratesremote application programs185 as residing onremote computer180. It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computers may be used.
• To further help understand the usefulness of the present invention, it may helpful to provide a brief description of a[0041]speech synthesizer300 illustrated in FIG. 3. However, it should be noted that thesynthesizer300 is provided for exemplary purposes and is not intended to limit the present invention.
• FIG. 2 is a block diagram of a[0042]mobile device200, which is an exemplary computing environment.Mobile device200 includes amicroprocessor202,memory204, input/output (I/O)components206, and acommunication interface208 for communicating with remote computers or other mobile devices. In one embodiment, the afore-mentioned components are coupled for communication with one another over asuitable bus210.
• [0043]Memory204 is implemented as non-volatile electronic memory such as random access memory (RAM) with a battery back-up module (not shown) such that information stored inmemory204 is not lost when the general power tomobile device200 is shut down. A portion ofmemory204 is preferably allocated as addressable memory for program execution, while another portion ofmemory204 is preferably used for storage, such as to simulate storage on a disk drive.
• [0044]Memory204 includes anoperating system212,application programs214 as well as anobject store216. During operation,operating system212 is preferably executed byprocessor202 frommemory204.Operating system212, in one preferred embodiment, is a WINDOWS® CE brand operating system commercially available from Microsoft Corporation.Operating system212 is preferably designed for mobile devices, and implements database features that can be utilized byapplications214 through a set of exposed application programming interfaces and methods. The objects inobject store216 are maintained byapplications214 andoperating system212, at least partially in response to calls to the exposed application programming interfaces and methods.
• [0045]Communication interface208 represents numerous devices and technologies that allowmobile device200 to send and receive information. The devices include wired and wireless modems, satellite receivers and broadcast tuners to name a few.Mobile device200 can also be directly connected to a computer to exchange data therewith. In such cases,communication interface208 can be an infrared transceiver or a serial or parallel communication connection, all of which are capable of transmitting streaming information.
• Input/[0046]output components206 include a variety of input devices such as a touch-sensitive screen, buttons, rollers, and a microphone as well as a variety of output devices including an audio generator, a vibrating device, and a display. The devices listed above are by way of example and need not all be present onmobile device200. In addition, other input/output devices may be attached to or found withmobile device200 within the scope of the present invention.
• Under the present invention, a speech synthesizer is provided that concatenates stored samples of speech units without modifying the prosody of the samples. The present invention is able to achieve a high level of naturalness in synthesized speech with a carefully designed speech corpus by storing samples based on the prosodic and phonetic context in which they occur. In particular, the present invention limits the training text to those sentences that will produce the most frequent sets of prosodic contexts for each speech unit. The present invention also provides a multi-tier selection mechanism for selecting a set of samples that will produce the most natural sounding speech.[0047]
• FIG. 3 is a block diagram of a[0048]speech synthesizer300 that is capable of constructing synthesizedspeech302 from aninput text304 under embodiments of the present invention. In conventional concatenative TTS systems, a pitch and duration modification algorithm, such as PSOLA, is applied to pre-stored units to guarantee that the prosodic features of synthetic speech meet the predicted target values. These systems have the advantages of flexibility in controlling the prosody. Yet, they often suffer from significant quality decrease in naturalness. In theTTS system300, speech is generated by directly concatenating speech segments (for speech units such as syllables, phonemes, diphones, semiphones, etc.) without any pitch or duration modification under the assumption that the speech database contains enough prosodic and spectral varieties for all speech units and the best fitting segments can always be found.
• Before[0049]speech synthesizer300 can be utilized to constructspeech302, it must be initialized with samples of speech units taken from atraining text306 that is read intospeech synthesizer300 astraining speech308.
• As noted above, speech synthesizers are constrained by a limited size memory. Because of this,[0050]training text306 must be limited in size to fit within the memory. However, if the training text is too small, there will not be enough samples of the training speech to allow for concatenative synthesis without prosody modifications. One aspect of the present invention overcomes this problem by trying to identify a set of speech units in a very large text corpus that must be included in the training text to allow for concatenative synthesis without prosody modifications.
• FIG. 4 provides a block diagram of components used to identify[0051]smaller training text306 of FIG. 3 from a verylarge corpus400. Under one embodiment, verylarge corpus400 is a corpus of five years worth of the People's Daily, a Chinese newspaper, and contains about 97 million Chinese Characters.
• Initially,[0052]training text400 is parsed by a parser/semantic identifier402 into strings of individual speech units attached with various textual information. Under some embodiments of the invention, especially those used to form Chinese speech, the speech units are tonal syllables. However, other speech units such as phonemes, diphones, triphones or the mix of them may be used within the scope of the present invention.
• Parser/[0053]semantic identifier402 also identifies high-level prosodic information about each sentence provided to theparser402. This high-level prosodic information includes the predicted tonal levels for each speech unit as well as the grouping of speech units into prosodic words and phrases. In embodiments where tonal syllable speech units are used, parser/semantic identifier402 also identifies the first and last phoneme in each speech unit.
• The strings of speech units attached with textual and prosodic information produced from the training text are provided to a[0054]context vector generator404, which generates a Speech-unit Dependent Descriptive Contextual Variation Vector (SDDCVV, hereinafter referred to as a “context vector”). The context vector describes several context variables that can affect the naturalness of the speech unit. Under one embodiment, the context vector describes six variables or coordinates of textual information. They are:
• Position in phrase (PinP): the position of the current speech unit in its carrying prosodic phrase.[0055]
• Position in word (PinW): the position of the current speech unit in its carrying prosodic word.[0056]
• Left phonetic context (LPhC): category of the last phoneme in the speech unit to the left (preceding) of the current speech unit.[0057]
• Right phonetic context (RPhC): category of the first phoneme in the speech unit to the right (following) of the current speech unit.[0058]
• Left tone context (LTC): the tone category of the speech unit to the left (preceding) of the current speech unit.[0059]
• Right tone context (RTC): the tone category of the speech unit to the right (following) of the current speech unit.[0060]
• If desired, the coordinates of the context vector can also include the stress level of the current speech unit, the tonal identity of current speech unit or the coupling degree of its pitch, duration and/or energy with its neighboring units.[0061]
• Under one embodiment, the position in phrase coordinate and the position in word coordinate can each have one of four values, the left phonetic context can have one of eleven values, the right phonetic context can have one of twenty-six values and the left and right tonal contexts can each have one of two values. Under this embodiment, there are 4*4*11*26*2*2=18304 possible context vectors for each speech unit.[0062]
• The context vectors produced by[0063]generator404 are grouped based on their speech unit. For each speech unit, a frequency-basedsorter406 identifies the most frequent context vectors for each speech unit. The most frequently occurring context vectors for each speech unit are then stored in a list ofnecessary context vectors408. In one embodiment, the top context vectors, whose accumulated frequency of occurrence is not less than half of the total frequency of occurrence of all units, are stored in the list.
• The sorting and pruning performed by[0064]sorter406 is based on a discovery made by the present inventors. In particular, the present inventors have found that certain context vectors occur repeatedly in the corpus. By making sure that these context vectors are found in the training corpus, the present invention increases the chances of having an exact context match for an input text without greatly increasing the size of the training corpus. For example, the present inventors have found that by ensuring that the top two percent of the context vectors are represented in the training corpus, an exact context match will be found for an input text speech unit over fifty percent of the time.
• Using the list of[0065]necessary context vectors408, atext selection unit410 selects sentences from verylarge corpus400 to producetraining text subset306. In a particular embodiment,text selection unit410 uses a greedy algorithm to select sentences fromcorpus400. Under this greedy algorithm,selection unit410 scans all sentences in the corpus and picks out one at a time to add to the selected group.
• During the scan,[0066]selection unit410 determines how many context vectors inlist408 are found in each sentence. The sentence that contains the maximum number of needed context vectors is then added totraining text306. The context vectors that the sentence contains are removed fromlist408 and the sentence is removed from thelarge text corpus400. The scanning is repeated until all of the context vectors have been removed fromlist408.
• After[0067]training text subset306 has been formed, it is read by a person and digitized into a training speech corpus. Both the training text and training speech can be used to initializespeech synthesizer300 of FIG. 3. This initialization begins by parsing the sentences oftext306 into individual speech units that are annotated with high-level prosodic information. In FIG. 3, this is accomplished by a parser/semantic identifier310, which is similar to parser/semantic identifier402 of FIG. 4. The parsed speech units and their high-level prosodic description are then provided to acontext vector generator312, which is similar tocontext vector generator404 of FIG. 4.
• The context vectors produced by[0068]context vector generator312 are provided to acomponent storing unit314 along with speech samples produced by asampler316 fromtraining speech signal308. Each sample provided bysampler316 corresponds to a speech unit identified byparser310.Component storing unit314 indexes each speech sample by its context vector to form an indexed set of storedspeech components318.
• Under one embodiment, the samples are indexed, for example, by a prosody-dependent decision tree (PDDT), which is formed automatically using a classification and regression tree (CART). CART provides a mechanism for selecting questions that can be used to divide the stored speech components into small groups of similar speech samples. Typically, each question is used to divide a group of speech components into two smaller groups. With each question, the components in the smaller groups become more homogenous. The process for using CART to form the decision tree is shown in FIG. 5.[0069]
• At[0070]step500 of FIG. 5, a list of candidate questions is generated for the decision tree. Under one embodiment, each question is directed toward some coordinate or combination of coordinates in the context vector.
• At[0071]step502, an expected square error is determined for all of the training samples fromsampler316. The expected square error gives a measure of the distances among a set of features of each sample in a group. In one particular embodiment, the features are prosodic features of average fundamental frequency (F_a), average duration (F_b), and range of the fundamental frequency (F_c) for a unit. For this embodiment, the expected square error is defined as:
• ESE(t)=E(W_aE_a+W_bE_b+W_cE_c)  EQ. 1
• where ESE(t) is the expected square error for all samples X on node t in the decision tree, E[0072]_a, E_b, and E_care the square error for F_a, F_b, and F_c, respectively, W_a, W_b, and W_care weights, and the operation of determining the expected value of the sum of square errors is indicated by the outer E( ).
• Each square error is then determined as:[0073]
• E_j=|F_j−R(F_j)|²=,j=a,b,c  EQ. 2
• where R(F[0074]_j) is a regression value calculated from samples X on node t. In this embodiment, the regression value is the expected value of the feature as calculated from the samples X at node t:
• R_j(F_j)=E(F_j/X∈node_t).
• Once the expected square error has been determined at[0075]step502, the first question in the question list is selected atstep504. The selected question is applied to the context vectors atstep506 to group the samples into candidate sub-nodes for the tree. The expected square error of each sub-node is then determined atstep508 usingequations 1 and 2 above.
• At[0076]step510, a reduction in expected square error created by generating the two sub-nodes is determined. Under one embodiment, this reduction is calculated as:
• ΔWESE(t)=ESE(t)P(t)−(ESE(l)P(l)+ESE(r)P(r))  EQ. 3
• where ΔWESE(t) is the reduction in expected square error, ESE(t) is the expected square error of node t, against which the question was applied, P(t) is the percentage of samples in node t, ESE(l) and ESE(r) are the expected square error of the left and right sub-nodes formed by the question, respectively, and P(l) and P(r) are the percentage of samples in the left and right node, respectively.[0077]
• The reduction in expected square error provided by the current question is stored and the CART process determines if the current question is the last question in the list at[0078]step512. If there are more questions in the list, the next question is selected at step514 and the process returns to step506 to divide the current node into sub-nodes based on the new question.
• After every question has been applied to the current node at[0079]step512, the reductions in expected square error provided by each question are compared and the question that provides the greatest reduction is set as the question for the current node of the decision tree atstep515.
• At[0080]step516, a decision is made as to whether or not the current set of leaf nodes should be further divided. This determination can be made based on the number of samples in each leaf node or the size of the reduction in square error possible with further division.
• Under one embodiment, when the decision tree is in its final form, each leaf node will contain a number of samples for a speech unit. These samples have slightly different prosody from each other. For example, they may have different phonetic contexts or different tonal contexts from each other. By maintaining these minor differences within a leaf node, this embodiment of the invention introduces slender diversity in prosody, which is helpful in removing monotonous prosody.[0081]
• If the current leaf nodes are to be further divided at[0082]step516, a leaf node is selected atstep518 and the process returns to step504 to find a question to associate with the selected node. If the decision tree is complete atstep516, the process of FIG. 5 ends atstep520.
• The process of FIG. 5 results in a prosody-[0083]dependent decision tree320 of FIG. 3 and a set of storedspeech samples318, indexed bydecision tree320. Once created,decision tree320 andspeech samples318 can be used under further aspects of the present invention to generate concatenative speech without requiring prosody modification.
• The process for forming concatenative speech begins by parsing a sentence in[0084]input text304 using parser/semantic identifier310 and identifying high-level prosodic information for each speech unit produced by the parse. This prosodic information is then provided tocontext vector generator312, which generates a context vector for each speech unit identified in the parse. The parsing and the production of the context vectors are performed in the same manner as was done during the training ofprosody decision tree320.
• The context vectors are provided to a[0085]component locator322, which uses the vectors to identify a set of samples for the sentence. Under one embodiment,component locator322 uses a multi-tier non-uniform unit selection algorithm to identify the samples from the context vectors.
• FIGS. 6 and 7 provide a block diagram and a flow diagram for the multi-tier non-uniform selection algorithm. In[0086]step700, each vector in the set of input context vectors is applied to prosody-dependent decision tree320 to identify aleaf node array600 that contains a leaf node for each context vector. Atstep702, a set of distances is determined by adistance calculator602 for each input context vector. In particular, a separate distance is calculated between the input context vector and each context vector found in its respective leaf node. Under one embodiment, each distance is calculated as:$\begin{array}{cc}{D}_c=\sum _{i=1}^IW_{\mathrm{ci}}D_i& \mathrm{EQ}.4\end{array}$

  
• where D[0087]_cis the context distance, D_iis the distance for coordinate i of the context vector, W_ciis a weight associated with coordinate i, and I is the number of coordinates in each context vector.
• At[0088]step704, the N samples with the closest context vectors are retained while the remaining samples are pruned fromnode array600 to form prunedleaf node array604. The number of samples, N, to leave in the pruned nodes is determined by balancing improvements in prosody with improved processing time. In general, more samples left in the pruned nodes means better prosody at the cost of longer processing time.
• At[0089]step706, the pruned array is provided to aViterbi decoder606, which identifies a lowest cost path through the pruned array. Under a single-tier embodiment of the present invention, the lowest cost path is identified simply by selecting the sample with the closest context vector in each node. Under a multi-tier embodiment, the cost function is modified to be:$\begin{array}{cc}{C}_c=W_c\sum _{j=1}^JD_{\mathrm{cj}}+W_s\sum _{j=1}^JC_{\mathrm{sj}}& \mathrm{EQ}.5\end{array}$

  
• where C[0090]_cis the concatenation cost for the entire sentence, W_cis a weight associated with the distance measure of the concatenated cost, D_cjis the distance calculated inequation 4 for the j^thspeech unit in the sentence, W_sis a weight associated with a smoothness measure of the concatenated cost, C_sjis a smoothness cost for the j^thspeech unit, and J is the number of speech units in the sentence.
• The smoothness cost in[0091]Equation 5 is defined to provide an objective measure of the prosodic mismatch between sample j and the samples proposed as the neighbors to sample j by the Viterbi decoder. Under one embodiment, the smoothness cost is determined based on whether a sample and its neighbors were found as neighbors in an utterance in the training corpus. If a sample occurred next to its neighbors in the training corpus, the smoothness cost is zero since the samples contain the proper spectral transition in between. If a sample did not occur next to its neighbors in the training corpus (referred as non-neighboring case), the smoothness cost is set to one. Under another embodiment, different values are assigned to the smoothness cost for non-neighboring cases according to their boundary types. For example, if the boundary between the two segments is sonorant to sonorant, the largest cost (1) is given. If the boundary between them is non-sonorant consonant to non-sonorant consonant, a small cost (0.2) is given. The cost between sonorant to non-sonorant or non-sonorant to sonorant transition is in middle (0.5). The different smoothness costs lead the search algorithm to prefer concatenation at boundaries with smaller cost.
• Using the multi-tier non-uniform approach, if a large block of speech units, such as a word or a phrase, in the input text exists in the training corpus, preference will be given to selecting all of the samples associated with that block of speech units. Note, however, that if the block of speech units occurred within a different prosodic context, the distance between the context vectors will likely cause different samples to be selected than those associated with the block.[0092]
• Once the lowest cost path has been identified by[0093]Viterbi decoder606, the identifiedsamples608 are provided tospeech constructor303. With the exception of small amounts of smoothing at the boundaries between the speech units,speech constructor303 simply concatenates the speech units to form synthesizedspeech302. Thus, the speech units are combined without having to change their prosody.
• The cost function or objective measure provided above contains only first order components of the seven textual factors, yet, higher order interactions might exist among these factors. In further embodiments, the context vector, cost function or objective measure comprises one or more higher order coordinates being combinations of at least two factors from a set of factors including: an indication of a position of a speech unit in a phrase; an indication of a position of a speech unit in a word; an indication of a category for a phoneme preceding a speech unit; an indication of a category for a phoneme following a speech unit; an indication of a category for tonal identity of the current speech unit; an indication of a category for tonal identity of a preceding speech unit; an indication of a category for tonal identity of a following speech unit; an indication of a level of stress of a speech unit; an indication of a coupling degree of pitch, duration and/or energy with a neighboring unit; and an indication of a degree of spectral mismatch with a neighboring speech unit. A technique for obtaining an optimized cost function, which can be used in the[0094]speech synthesizer300 as the criterion for the unit selection, is also provided below.
• Evaluating the quality of synthesized speech contains two aspects, intelligibility and naturalness. Generally, intelligibility is not a large concern for most text-to-speech systems. However, the naturalness of synthesized speech is a larger issue and is still far from most expectations.[0095]
• During text-to-speech system development, it is necessary to have regular evaluations on a naturalness of the system. The Mean Opinion Score (MOS) is one of the most popular and widely accepted subjective measures for naturalness. However, running a formal MOS evaluation is expensive and time consuming. Generally, to obtain a MOS score for a system under consideration, a collection of synthesized waveforms must be obtained from the system. The synthesized waveforms, together with some waveforms generated from other text-to-speech systems and/or waveforms uttered by a professional announcer are randomly played to a set of subjects. Each of the subjects are asked to score the naturalness of each waveform from 1-5 (1=bad, 2=poor, 3=fair, 4=good, 5=excellent). The means of the scores from the set of subjects for a given waveform represents naturalness in a MOS evaluation. Recently, a method for estimating mean opinion score or naturalness of synthesized speech has been advanced by Chu, M. and Peng, H., in “An objective measure for estimating MOS of synthesized speech”,[0096]Proceedings of Eurospeech2001, 2001, and is discussed further below. The method includes using an objective measure that has components derived directly from textual information used to form synthesized utterances. The objective measure has a high correlation with the mean opinion score such that a relationship can be formed between the objective measure and the corresponding mean opinion score. An estimated mean opinion score can be obtained easily from the relationship when the objective measure is applied to utterances of a modified speech synthesizer.
• The objective measure can be based on one or more factors of the speech units used to create the utterances. The factors can include the position of the speech unit in a phrase or word, the neighboring phonetic or tonal context, the spectral mismatch of successive speech units or the stress level of the speech unit. Weighting factors can be used since correlation of the factors with mean opinion score has been found to vary between the factors.[0097]
• By using the objective measure it is easy to track performance in naturalness of the speech synthesizer, thereby allowing efficient development of the speech synthesizer. In particular, the objective measure can serve as criteria for optimizing the algorithms for speech unit selection and speech database pruning.[0098]
• As discussed above, the evaluation of concatenative cost can form the basis of an objective measure for MOS estimation. A method for using the objective measure in estimating MOS is illustrated in FIG. 8. Generally, the method includes generating a set of synthesized utterances at[0099]step800, and subjectively rating each of the utterances atstep802. A score is then calculated for each of the synthesized utterances using the objective measure atstep804. The scores from the objective measure and the ratings from the subjective analysis are then analyzed to determine a relationship atstep806. The relationship is used atstep808 to estimate naturalness or MOS when the objective measure is applied to the textual information of speech units for another utterance or second set of utterances from a modified speech synthesizer (e.g. when a parameter of the speech synthesizer has been changed). It should be noted that the words of the “another utterance” or the “second set of utterances” obtained from the modified speech synthesizer can be the same or different words used in the first set of utterances.
• In one embodiment, in order to make the concatenative cost comparable among utterances with variable number of syllables, the average concatenative cost of an utterance is used and can be expressed as:[0100]$\begin{array}{cc}{C}_a=\sum _{i=1}^{I+1}W_iC_{\mathrm{ai}}C_{\mathrm{ai}}=\{\begin{array}{cc}\frac{1}{J}\sum _{l=1}^JD_i\left(l\right),& i=1,\cdots ,I\\ \frac{1}{J-1}\sum _{l=1}^{J-1}C_s\left(l\right),& i=I+1\end{array}W_i=\{\begin{array}{cc}{W}_{\mathrm{ci}}W_c& i=1,\cdots ,I\\ W_s& i=I+1\end{array}& \mathrm{EQ}.6\end{array}$

  
• where, C[0101]_ais the average concatenative cost and C_ai(i=1, . . . ,7) one or more of the factors that contribute to C_a, which are, in the illustrative embodiment, the average costs for position in phrase (“PinP”), position in word (“PinW”), left phonetic context (“LPhC”), right phonetic context (“RPhC”), left tone context (“LTC”), right tone context (“RTC”) and smoothness cost per unit in an utterance.
• The cost function as provided above is a weighted sum of seven factors. Six of the factors are distances between the target category and the category of candidate unit (named as unit category) for the six contextual factors, which are PinP, PinW, LPhC, RPhC, LTC and RTC. Since all these factors take only categorical values, the distance between categories are empirically predefined in distance tables. The seventh factor is an enumerated smoothness cost, which takes value 0 when current candidate unit is a continuous segment with the unit before it in the unit inventory and takes value larger than 0 otherwise.[0102]
• W[0103]_iare weights for the seven component-costs and all are set to 1, but can be changed. For instance, it has been found that the coordinate having the highest correlation with mean opinion score was smoothness, whereas the lowest correlation with mean opinion score was position in phase. It is therefore reasonable to assign larger weights for components with high correlation and smaller weights for components with low correlation. In one experiment, the following weights were used:
• Position in Phrase, W[0104]₁=0.10
• Position in Word, W[0105]₂=0.60
• Left Phonetic Context, W[0106]₃=0.10
• Right Phonetic Context, W[0107]₄=0.76
• Left Tone Context, W[0108]₅=1.76
• Right Tone Context, W[0109]₆=0.72
• Smoothness, W[0110]₇=2.96
• In one exemplary embodiment, 100 sentences are carefully selected from a 200 MB text corpus so the C[0111]_aand C_ai(i=1, . . . ,7) of them are scattered into wide spans. Four synthesized waveforms are generated for each sentence with thespeech synthesizer200 above with four speech databases, whose sizes are 1.36 GB, 0.9 GB, 0.38 GB and 0.1 GB, respectively. C_aand C_aiof each waveform are calculated. All the 400 synthesized waveforms, together with some waveforms generated from other TTS systems and waveforms uttered by a professional announcer, are randomly played to 30 subjects. Each of the subjects is asked to score the naturalness of each waveform from 1-5 (1=bad, 2=poor, 3=fair, 4=good, 5=excellent). The mean of the thirty scores for a given waveform represents its naturalness in MOS.
• Fifty original waveforms uttered by the speaker who provides voice for the speech database are used in this example. The average MOS for these waveforms was 4.54, which provides an upper bound for MOS of synthetic voice. Providing subjects a wide range of speech quality by adding waveforms from other systems can be helpful so that the subjects make good judgements on naturalness. However, only the MOS for the 400 waveforms generated by the speech synthesizer under evaluation are used in conjunction with the corresponding average concatenative cost score.[0112]
• FIG. 9 is a plot illustrating the objective measure (average concatenative cost) versus subjective measure (MOS) for the 400 waveforms. A correlation coefficient between the two dimensions is −0.822, which reveals that the average concatenative cost function replicates, to a great extent, the perceptual behavior of human beings. The minus sign of the coefficient means that the two dimensions are negatively correlated. The larger C[0113]_ais, the smaller the corresponding MOS will be. Alinear regression trendline902 is illustrated in FIG. 9 and is estimated by calculating the least squares fit throughout points. The trendline or curve is denoted as the average concatenative cost-MOS curve and for the exemplary embodiment is:
• Y=−1.0327x+4.0317.
• However, it should be noted that analysis of the relationship of average concatenative cost and MOS score for the representative waveforms can also be performed with other curve-fitting techniques, using, for example, higher-order polynomial functions. Likewise, other techniques of correlating average concatenative cost and MOS can be used. For instance, neural networks and decision trees can also be used.[0114]
• Using the average concatenative cost vs. MOS relationship, an estimate of MOS for a single synthesized speech waveform can be obtained by its average concatenative cost. Likewise, an estimate of the average MOS for a TTS system can be obtained from the average of the average of the concatenative costs that are calculated over a large amount of synthesized speech waveforms. In fact, when calculating the average concatenative cost, it is unnecessary to generate the speech waveforms since the costs can be calculated after the speech units have been selected.[0115]
• Although the concatenative cost function has proven to replicate, to a great extent, the perceptual behavior of human beings, it might not be optimal. It has been discovered by the inventors that some factors that can contribute to inaccuracies in the concatenative cost function include that many parameters in the cost function are assigned empirically by a human expert, and accordingly, they might not be the most suitable values. In addition, the concatenative cost function provided above contains only first order components of the seven textual factors, yet, higher order interactions might exist among these factors. Furthermore, there might be other components that could be added into the concatenative cost function.[0116]
• A method for optimizing the objective measure or concatenative cost function for unit selection in the corpus-based TTS system by maximizing the correlation between the concatenative cost and the MOS is provided. The method is illustrated in FIG. 10 at[0117]1000. Atstep1002, a subjective evaluation should be done first as discussed above. However, a beneficial aspect of this step is to log or record in a file or other means the textual information of all units appearing in the synthetic utterances evaluated. Atstep1003, an initial concatenative cost function is used and a correlation with MOS is established. Atstep1004, the concatenative cost function is altered, for example, using any one or more of the techniques described below. With the recorded log file, a new concatenative cost can be recalculated atstep1006 using the new a cost function. The correlation between the new concatenative cost and MOS is obtained atstep1008, which also serves as a measure for the validity of any change in the concatenative cost function.Steps1004,1006 and1008 can be repeated as necessary until an optimized concatenative cost function is realized. It is important to note that only a single run of MOS evaluation (step1002) is required in theoptimization method1000. This is helpful becausestep1002 can be particularly labor and time consuming. Other optimization algorithms such as Gradient Declination can also be used to optimize the free parameters.
• As indicated above, in order to evaluate improvements and accuracy made to the cost function, a measure needs to be used. One useful measure has been found to be the correlation between the concatenative cost and the MOS such as illustrated in FIG. 9. Thus, if the correlation between concatenative cost and MOS improves with changes to the concatenative cost function, such changes can be included in the concatenative cost function.[0118]
• As mentioned above, the log file of[0119]step1002 keeps the information of the target units wanted and the units actually used. Concatenative cost for all sentences can be calculated with any new cost function from the log file. That is to say, the form of the cost function or the distance tables used by the cost function can be changed, and the validity of the change can be measured through movement of the correlation between the new cost and the MOS for the set of synthesized utterances. Furthermore, when a specific format is given to a cost function, the correlation between concatenative cost and MOS can be treated as a function of the parameters of the concatenative cost function, denoted by the following equation
• Corr=f(x₁, x₂, . . . , x_N)  EQ. 7
• where, N is number of free parameters in the concatenative cost function. If the concatenative cost function is defined as equation (3), distances between target and unit categories for the six textual factors and the weights for the seven factors can be free parameters. An optimization routine is used to optimize the free parameters so that the largest correlation is to be achieved. One suitable optimization routine that can be used is the function “fmincon” in the Matlab Optimization Toolbox by The MathWorks, Inc. of Natick, Mass., U.S.A (“Optimization Toolbox User's Guide: For Use with MATLAB”), which searches for the minimum of a constrained nonlinear multivariable function, and optimizes the free parameters so that the largest correlation is to be achieved. Since concatenative cost and MOS is negatively correlated, Corr in[0120]equation 4 is to be minimized.
• In one embodiment, for instance, depending on the number of utterances available with MOS scores, the number of free parameters in each run of optimization should not be too large. Thus, in one embodiment, optimization can be separated into many runs. In each run at[0121]step1004, only some of the parameters are optimized and the others are fixed at their original values. Referring to FIG. 11, three different kinds of changes can be made to the concatenative cost function; specifically optimize the distance tables for the six single-order textual factors individually atstep1102; explore the interactions among factors and add some higher order components into the cost function atstep1104; and optimize the weight for each component in the new cost function atstep1106.
• Since some parameters in the distance tables may not be used frequently depending on the number of available sentences, optimizing them with a few observations will probably cause an overfitting problem. In this case, to avoid overfitting, a threshold can be set for the number of times a parameter had been used. Only frequently used ones are optimized. Though, no globally optimized solution is guaranteed, it is quite likely that the overall correlation is increased.[0122]
• In order to check the validity of the optimized parameters, a K-fold cross validation experiment is done. In one embodiment, K is set to 4. In each run of optimization, only 300 utterances are used for training and the remaining[0123]100 sentences are used for testing. If the difference between average correlation coefficients for the training and testing set is large, the optimization is considered invalid. Thus, the number of free parameters should be reduced. For valid optimization, means of the four sets of optimized parameters are used in the final cost function.
• As indicated above, the distances between target categories and unit categories of a textual factor are assigned manually, which may not be the most suitable values. In a first method of optimization of the concatenative cost function, the distance table for each textual factor is improved at[0124]step1102 individually. Here, the concatenative cost function contains only a single textual component in each run of optimization. The correlation coefficients between the six textual factors and MOS before and after optimization are listed in Table 1.

  	TABLE 1
  		IniCorr	TrCorr	TsCorr

  	PinP	0.498	0.525	0.498
  	PinW	0.623	0.631	0.623
  	LPhC	0.553	0.715	0.703
  	RPhC	0.688	0.742	0.736
  	LTC	0.654	0.743	0.731
  	RTC	0.622	0.755	0.732

• In Table 1, the correlation coefficients between the six textual factors and MOS before and after optimization are provided where “IniCorr” provides the initial coefficient obtained with the empirical distance tables; “TrCorr” provides the average coefficient on the four training sets after optimization; and “TsCorr” provides the average coefficient on testing sets after optimization.[0125]
• It can be seen that there is no change for the correlation for the factor PinP and PinW on the testing set, and both of them have smaller correlation to MOS than other factors. The reason might be that both of them have been used in the splitting question for constructing indexing CART for the unit inventory. Thus, most of the units used in subjective experiment have zero distances for the two factors. For the other four factors, great increases are obtained.[0126]
• Using the factor RTC by way of example for detailed explanation, the initial distance table and the optimized one for RTC are given in Table 2(a) and 2(b) below. T1-T5 represent the four normal tones and the neutral tone in Mandarin Chinese. Rows in Tables 2(a) and 2(b) represent the target RTC, while the columns represent the unit RTC. The numbers in the tables are the distances between target RTC and unit RTC. It can be seen that many distances reach a more precise value after optimization, in comparison to those given by a human expert. There are some numbers unchanged in Table 2(b) since they haven't been used enough times in the training set. Thus, they are fixed at the initial values during the optimizing phase.[0127]

  TABLE 2(a)
  The initial distance table

  Unit

  Target RTC

  	RTC	T1	T2	T3	T4	T5

  	Ti	0	0.25	0.75	0.25	1
  	T2	0.5	0	0.25	0.75	1
  	T3	0.5	0.25	0	0.75	1
  	T4	0.75	0.5	1	0	0.25
  	T5	0.5	0.75	1	0.25	0

• [0128]

  TABLE 2(b)
  The optimized distance table

  Unit

  Target RTC

  	RTC	T1	T2	T3	T4	T5

  	T1	0	0.25	0.75	0.93	0.27
  	T2	0.62	0	0.37	0.95	1
  	T3	0.5	0.88	0	0.75	1
  	T4	0.87	0.56	1	0	0.58
  	T5	0.5	0.75	1	0.66	0

• As provided above in equation 6, the concatenative cost function is a linear combination of the seven factors. Yet, it has been discovered some of them may have interactions. However, to limit the number of free parameters, the numbers of categories for the six textual factors can be reduced, if desired. In the discussion provided below, the number of categories for PinP and PinW have been reduced to 2, while LPhC have been reduced to 4 and RPhC, LTC and RTC have been reduced to 3, although this should not be considered necessary or limiting. In the exemplary optimization method discussed herein, six second-order combinations (i.e. combinations of two textual factors) are investigated at[0129]step1104, in which the maximum number of free parameters is 36; however it should be understood other combinations and/or even higher order combinations (combinations of three or more textual factors) can also be used. In the present discussion, the combination between LPhC and other factors has not been adopted since these combinations may cause too many free parameters.
• As with the single-order components of the cost function, the higher-order components also take only categorical values, the distance between categories are empirically predefined in distance tables. After optimizing the distance tables for these second-order components individually in a manner similar to that discussed above with the single-order textual factors in[0130]step1104, their correlation coefficients to MOS are listed in Table 3. Comparing Table 3 to Table 1, it can be seen that all combinations of Table 3 have a higher correlation than using PinP and PinW alone, yet, only coefficients for LTC-PinW and LTC-PinW pairs are higher than those of using LTC alone. It appears that some of the second-order components play important roles for unit selection. In a further embodiment discussed below, all of the higher-order components are used to form the concatenative cost function at first and some of them are then removed after optimizing the weights since they receive small weights.

  	TABLE 3
  		RPhC	LTC	RTC

  	PinP	0.719	0.752	0.710
  	PinW	0.751	0.790	0.745

• An enumerated smoothness cost is used in the original cost function. Various smoothness costs based on the combinations with the six textual factors have been investigated. In one embodiment, it has been found beneficial to assign the smoothness cost by considering PinW (2 categories), LTC (3 categories) and the final type of current unit (3 categories). That is to say, when the current unit is a continuous segment of its previous segment in the unit inventory, its smoothness cost is set to be zero, otherwise, it is to be assigned a value from a table of 18(=2*3*3) possibilities according the conditions described above. The values in the smoothness cost table can be optimized by maximizing the correlation between smoothness cost and MOS in the training sets, e.g. four training sets. After optimization, the correlation coefficient reaches 0.883, which is higher than the old one, 0.846. This reveals that the new smoothness cost is more suitable than the original one and is used to replace the original one in the cost function discussed below.[0131]
• Since it is not generally known which component is more important, at first, the new cost function in[0132]step806 is formed by weighted sum of all the single-order components and higher-order components as discussed above. The weights for each of the components are then optimized as discussed above. An example of optimized weights for 13 components is provided in Table 4. Since some of the components received very small weights, they can be removed from the cost function without much effect. In a further embodiment,step808 includes removing some components below a selected threshold and keeping only the more significant components (identified with stars), wherein the weights of the remaining components are optimized again. The new optimized weights in the final concatenative cost function for seven components are given in Table 5. The correlation coefficient between the final cost and MOS reaches 0.897, which is much higher than the original one, 0.822. Speech synthesized with the new cost function should sound more natural than that generated with the original one.

  	TABLE 4
  	Component	Weight	Component	Weight

  	PinP	0.008	RPhC-PinP pair	0.008
  	PinW	0.008	RPhC-PinW pair	0.023
  	LPhC*	0.099	LTC-PinP pair*	0.088
  	RPhC*	0.054	LTC-PinW pair*	0.113
  	LTC*	0.104	RTC-PinP pair	0.008
  	RTC*	0.091	RTC-PinW pair	0.016
  			New smooth cost*	0.380

• [0133]

  	TABLE 5
  	Component	Weight	Component	Weight

  	LPhC	0.061	LTC-PinP pair	0.122
  	RPhC	0.059	LTC-PinW pair	0.170
  	LTC	0.016	New smooth cost	0.481
  	RTC	0.091

• At this point it should be noted the utterances used for the MOS experiment should be designed carefully so that units have wide coverage for textual factors. In the example discussed above, prosodic feature orientated CART indices have been adopted for all units, where most of the units used in the MOS evaluation take zero costs for their PinP and PinW factors. Thus, the two factors show smaller correlations to MOS, though they can be important factors. On the other hand, optimization using 400 utterances is not enough for training all the parameters. If possible, a larger scale MOS evaluation can be used to get more reliable optimized parameters. Since the result of MOS evaluation can be used perpetually, (i.e. over and over), it may be worthwhile to do a well-designed large-scale MOS evaluation.[0134]
• Although the present invention has been described with reference to particular embodiments, workers skilled in the art will recognize that changes may be made in form and detail without departing from the spirit and scope of the invention. In particular, although context vectors are discussed above, other representations of the context information sets may be used within the scope of the present invention.[0135]

Claims (30)

What is claimed is:

1. A method for synthesizing speech, the method comprising:

generating a training context vector for each of a set of training speech units in a training speech corpus, each training context vector indicating the prosodic context of a training speech unit in the training speech corpus;

indexing a set of speech segments associated with a set of training speech units based on the context vectors for the training speech units;

generating an input context vector for each of a set of input speech units in an input text, each input context vector indicating the prosodic context of an input speech unit in the input text;

using the input context vectors to find a speech segment for each input speech unit; and

concatenating the found speech segments to form a synthesized speech signal.

2. The method ofclaim 1 wherein the each context vector comprises a position-in-phrase coordinate indicating the position of the speech unit in a phrase.

3. The method ofclaim 1 wherein the each context vector comprises a position-in-word coordinate indicating the position of the speech unit in a word.

4. The method ofclaim 1 wherein the each context vector comprises a left phonetic coordinate indicating a category for the phoneme to the left of the speech unit.

5. The method ofclaim 1 wherein the each context vector comprises a right phonetic coordinate indicating a category for the phoneme to the right of the speech unit.

6. The method ofclaim 1 wherein the each context vector comprises a left tonal coordinate indicating a category for the tone of the speech unit to the left of the speech unit.

7. The method ofclaim 1 wherein the each context vector comprises a right tonal coordinate indicating a category for the tone of the speech unit to the right of the speech unit.

8. The method ofclaim 1 wherein the each context vector comprises a coordinate indicating a coupling degree of pitch, duration and/or energy with a neighboring unit.

9. The method ofclaim 1 the each context vector comprises a coordinate indicating a level of stress of a speech unit.

10. The method ofclaim 1 wherein indexing a set of speech segments comprises generating a decision tree based on the training context vectors.

11. The method ofclaim 10 wherein using the input context vectors to find a speech segment comprises searching the decision tree using the input context vector.

12. The method ofclaim 11 wherein searching the decision tree comprises:

identifying a leaf in the tree for each input context vector, each leaf comprising at least one candidate speech segments; and

selecting one candidate speech segment in each leaf node, wherein if there is more than one candidate speech segment on the node The selection is based on a cost function.

13. The method ofclaim 12 wherein the cost function comprises a distance between the input context vector and a training context vector associated with a speech segment.

14. The method ofclaim 13 wherein the cost function further comprises a smoothness cost that is based on a candidate speech segment of at least one neighboring speech unit.

15. The method ofclaim 14 wherein the smoothness cost gives preference to selecting a series of speech segments for a series of input context vectors if the series of speech segments occurred in series in the training speech corpus.

16. The method ofclaim 1 wherein the context vector comprises one or more higher order coordinates being combinations of at least two factors from a set of factors including:

an indication of a position of a speech unit in a phrase;

an indication of a position of a speech unit in a word;

an indication of a category for a phoneme preceding a speech unit;

an indication of a category for a phoneme following a speech unit;

an indication of a category for tonal identity of the current speech unit;

an indication of a category for tonal identity of a preceding speech unit;

an indication of a category for tonal identity of a following speech unit;

an indication of a level of stress of a speech unit;

an indication of a coupling degree of pitch, duration and/or energy with a neighboring unit; and

an indication of a degree of spectral mismatch with a neighboring speech unit.

17. A method of selecting sentences for reading into a training speech corpus used in speech synthesis, the method comprising:

identifying a set of prosodic context information for each of a set of speech units;

determining a frequency of occurrence for each distinct context vector that appears in a very large text corpus;

using the frequency of occurrence of the context vectors to identify a list of necessary context vectors; and

selecting sentences in the large text corpus for reading into the training speech corpus, each selected sentence containing at least one necessary context vector.

18. The method ofclaim 17 wherein identifying a collection of prosodic context information sets as necessary context information sets comprises:

determining the frequency of occurrence of each prosodic context information set across a very large text corpus; and

identifying a collection of prosodic context information sets as necessary context information sets based on their frequency of occurrence.

19. The method ofclaim 18 wherein identifying a collection of prosodic context information sets as necessary context information sets further comprises:

sorting the context information sets by their frequency of occurrence in decreasing order;

determining a threshold, F, for accumulative frequency of top context vectors; and

selecting the top context vectors whose accumulative frequency is not smaller than F for each speech unit as necessary prosodic context information sets.

20. The method ofclaim 17 further comprising indexing only those speech segments that are associated with sentences in the smaller training text and wherein indexing comprises indexing using a decision tree.

21. The method ofclaim 20 wherein indexing further comprises indexing the speech segments in the decision tree based on information in the context information sets.

22. The method ofclaim 21 wherein the decision tree comprises leaf nodes and at least one leaf node comprises at least two speech segments for the same speech unit.

23. A method of selecting speech segments for concatenative speech synthesis, the method comprising:

parsing an input text into speech units;

identifying context information for each speech unit based on its location in the input text and at least one neighboring speech unit;

identifying a set of candidate speech segments for each speech unit based on the context information; and

identifying a sequence of speech segments from the candidate speech segments based in part on a smoothness cost between the speech segments.

24. The method ofclaim 23 wherein identifying a set of candidate speech segments for a speech unit comprises applying the context information for a speech unit to a decision tree to identify a leaf node containing candidate speech segments for the speech unit.

25. The method ofclaim 24 wherein identifying a set of candidate speech segments further comprises pruning some speech segments from a leaf node based on differences between the context information of the speech unit from the input text and context information associated with the speech segments.

26. The method ofclaim 23 wherein identifying a sequence of speech segments comprises using a smoothness cost that is based on whether two neighboring candidate speech segments appeared next to each other in a training corpus.

27. The method ofclaim 23 wherein identifying a sequence of speech segments comprises using an objective measure comprising one or more first order components from a set of factors comprising:

an indication of a position of a speech unit in a phrase;

an indication of a position of a speech unit in a word;

an indication of a category for a phoneme preceding a speech unit;

an indication of a category for a phoneme following a speech unit;

an indication of a category for tonal identity of the current speech unit;

an indication of a category for tonal identity of a preceding speech unit;

an indication of a category for tonal identity of a following speech unit;

an indication of a level of stress of a speech unit;

an indication of a coupling degree of pitch, duration and/or energy with a neighboring unit; and

an indication of a degree of spectral mismatch with a neighboring speech unit.

28. The method ofclaim 23 wherein identifying a sequence of speech segments comprises using an objective measure comprising one or more higher order components being combinations of at least two factors from a set of factors including:

an indication of a position of a speech unit in a phrase;

an indication of a position of a speech unit in a word;

an indication of a category for a phoneme preceding a speech unit;

an indication of a category for a phoneme following a speech unit;

an indication of a category for tonal identity of the current speech unit;

an indication of a category for tonal identity of a preceding speech unit;

an indication of a category for tonal identity of a following speech unit;

an indication of a level of stress of a speech unit;

an indication of a coupling degree of pitch, duration and/or energy with a neighboring unit; and

an indication of a degree of spectral mismatch with a neighboring speech unit.

29. The method ofclaim 24 wherein identifying a sequence of speech segments further comprises identifying the sequence based in part on differences between context information for the speech unit of the input text and context information associated with a candidate speech segment.

30. A computer-readable medium having computer executable instructions for synthesizing speech from speech segments based on speech units found in an input text, the speech being synthesized through a method comprising steps of:

identifying context information for each speech unit based on the prosodic structure of the input text;

identifying a set of candidate speech segments for each speech unit based on the context information;

identifying a sequence of speech segments from the candidate speech segments;

concatenating the sequence of speech segments without modifying the prosody of the speech segments to form the synthesized speech.

Images