US20040204939A1 - Systems and methods for speaker change detection - Google Patents

Abstract

A speaker change detection system performs speaker change detection on an input audio stream. The speaker change detection system includes a segmentation component [401], a phone classification decode component [402], and a speaker change detection component [403]. The segmentation component [401] segments the audio stream into segments [501-504] of predetermined length intervals. The segments may overlap one another. The phone classification decode component decodes the intervals to produce a set of phone classes corresponding to each of the intervals. The speaker change detection component detects locations of speaker changes in the audio stream based on a similarity value calculated at phone class boundaries.

Description

RELATED APPLICATIONS
• This application claims priority under 35 U.S.C. § 119 based on U.S. Provisional Application No. 60/419,214 filed Oct. 17, 2002, the disclosure of which is incorporated herein by reference.[0001]
GOVERNMENT CONTRACT
• [0002] The U.S. Government may have a paid-up license in this invention and the right in limited circumstances to require the patent owner to license others on reasonable terms as provided for by the terms of Contract No. DARPA F30602-97-C-0253.
BACKGROUND OF THE INVENTION
• A. Field of the Invention[0003]
• The present invention relates generally to speech processing and, more particularly, to speaker change detection in an audio signal.[0004]
• B. Description of Related Art[0005]
• Speech has not traditionally been valued as an archival information source. As effective as the spoken word is for communicating, archiving spoken segments in a useful and easily retrievable manner has long been a difficult proposition. Although the act of recording audio is not difficult, automatically transcribing and indexing speech in an intelligent and useful manner can be difficult.[0006]
• Speech is typically received by a speech recognition system as a continuous stream of words without breaks. In order to effectively use the speech in information management systems (e.g., information retrieval, natural language processing and real-time alerting systems), the speech recognition system initially processes the speech to generate a formatted version of the speech. The speech may be transcribed and linguistic information, such as sentence structures, may be associated with the transcription. Additionally, information relating to the speakers may be associated with the transcription. Many speech recognition applications assume that all of the input speech originated from a single speaker. However, speech signals, such as news broadcasts, may include speech from a number of different speakers. It is often desirable to identify different speakers in the transcription of the broadcast. Before identifying an individual speaker, however, it is first necessary to determine when the different speakers begin and end speaking (called speaker change detection).[0007]
• FIG. 1 is a diagram illustrating speaker change detection using a conventional speaker change detection technique. An input audio stream is divided into a continuous stream of[0008]audio intervals101. The intervals each cover a fixed time duration such as 10 ms. Thus, in the example shown in FIG. 1, in which fiveintervals101 are shown, a total of 50 ms of audio is illustrated. Boundaries betweenintervals101, such as boundaries110-113, are considered to be potential candidates for a speaker change determination. The conventional speaker change detection system would examine all, or a predetermined number ofintervals101, to the left of a particular boundary, such asboundary112, and to the right ofboundary112. If the audio to the left and right ofboundary112 was determined to be dissimilar enough,boundary112 was assumed to be a speaker change. Whether the audio to the left and right ofboundary112 is similar or dissimilar may be based on comparing cepstral vectors for the audio.
• One drawback to the conventional speaker detection technique described above is that the interval boundaries occur at arbitrary locations. Accordingly, to ensure that a speaker change boundary is not missed, the intervals are assigned relatively short interval durations (e.g., 10 ms). Calculating whether audio segments are similar or dissimilar at each of the interval boundaries can, thus, be computationally burdensome.[0009]
• There is, therefore, a need in the art for improved speaker change detection techniques.[0010]
SUMMARY OF THE INVENTION
• Systems and methods consistent with the principles of this invention provide for fast speaker boundary change detection.[0011]
• A method consistent with aspects of the invention detects speaker changes in an input audio stream. The method includes segmenting the input audio stream into predetermined length intervals and decoding the intervals to produce a set of phones corresponding to each of the intervals. The method further includes generating a similarity measurement based on a first portion of the audio stream within an interval prior to a boundary between adjacent phones and a second portion of the audio stream within the interval and after the boundary. Further, the method includes detecting speaker changes based on the similarity measurement.[0012]
• A device consistent with another aspect of the invention detects speaker changes in an audio signal. The device comprises a processor and a memory containing instructions. The instructions, when executed by the processor, cause the processor to: segment the audio signal into predetermined length intervals and decode the intervals to produce a set of phones corresponding to the intervals. The instructions additionally cause the processor to generate a similarity measurement based on a first portion of the audio signal prior to a boundary between phones in one of the sets of phones and a second portion of the audio signal after the boundary, and detect speaker changes based on the similarity measurement.[0013]
• Yet another aspect of the invention is directed to a device for detecting speaker changes in an audio signal. The device comprises a segmentation component that segments the audio signal into predetermined length intervals and a phone classification decode component that decodes the intervals to produce a set of phone classes corresponding to each of the intervals. Further, a speaker change detection component detects locations of speaker changes in the audio signal based on a similarity value calculated over a first portion of the audio signal prior to a boundary between phone classes in one of the sets of phone classes and a second portion of the audio signal after the boundary in the one set of phone classes.[0014]
• Another aspect of the invention is directed to a system comprising an indexer, a memory system, and a server. The indexer receives input audio data and generates a rich transcription from the audio data. The rich transcription includes metadata that defines speaker changes in the audio data. The indexer further includes a segmentation component configured to divide the audio data into overlapping segments and a speaker change detection component. The speaker change detection component detects locations of speaker changes in the audio data based on a similarity value calculated at locations in the segments that correspond to phone class boundaries. The memory system stores the rich transcription and the server receives requests for documents and responds to the requests by transmitting ones of the rich transcriptions that match the requests.[0015]
BRIEF DESCRIPTION OF THE DRAWINGS
• The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate the invention and, together with the description, explain the invention. In the drawings,[0016]
• FIG. 1 is a diagram illustrating speaker change detection using a conventional speaker change detection technique;[0017]
• FIG. 2 is a diagram of a system in which systems and methods consistent with the present invention may be implemented;[0018]
• FIG. 3 is a diagram illustrating an exemplary computing device;[0019]
• FIG. 4 is diagram illustrating functional components of the speaker segmentation logic shown in FIG. 2;[0020]
• FIG. 5 is a diagram illustrating an audio stream broken into intervals consistent with an aspect of the invention;[0021]
• FIG. 6 is a diagram illustrating exemplary phone classes decoded for an audio interval; and[0022]
• FIG. 7 is a flow chart illustrating the operation of the speaker change detection component shown in FIG. 4.[0023]
DETAILED DESCRIPTION
• The following detailed description of the invention refers to the accompanying drawings. The same reference numbers in different drawings may identify the same or similar elements. Also, the following detailed description does not limit the invention. Instead, the scope of the invention is defined by the appended claims and equivalents.[0024]
• A speaker change detection (SCD) system consistent with the present invention performs speaker change detection on an input audio stream. The speaker change detection can be performed at real-time or faster than real-time. Speaker changes are detected only at phone boundaries detected within audio segments of a predetermined length. The predetermined length of the audio segments may be set at a much longer length, such as 30 seconds, than conventional predetermined intervals (e.g., 10 ms) used in detecting speaker changes.[0025]
Exemplary System
• FIG. 2 is a diagram of an[0026]exemplary system200 in which systems and methods consistent with the present invention may be implemented. In general,system200 provides indexing and retrieval of input audio forclients250. For example,system200 may index input speech data, create a structural summarization of the data, and provide tools for searching and browsing the stored data.
• [0027]System200 may includemultimedia sources210, anindexer220,memory system230, andserver240 connected toclients250 vianetwork260.Network260 may include any type of network, such as a local area network (LAN), a wide area network (WAN) (e.g., the Internet), a public telephone network (e.g., the Public Switched Telephone Network (PSTN)), a virtual private network (VPN), or a combination of networks. The various connections shown in FIG. 2 may be made via wired, wireless, and/or optical connections.
• [0028]Multimedia sources210 may include one or more audio sources, such as radio broadcasts, or video sources, such as television broadcasts.Indexer220 may receive audio data from one of these sources as an audio stream or file.
• Indexer[0029]220 may receive the input audio data frommultimedia sources210 and generate a rich transcription therefrom. For example,indexer220 may segment the input data by speaker, cluster audio segments from the same speaker, identify speakers by name or gender, and transcribe the spoken words.Indexer220 may also segment the input data based on topic and locate the names of people, places, and organizations.Indexer220 may further analyze the input data to identify when each word was spoken (possibly based on a time value).Indexer220 may include any or all of this information as metadata associated with the transcription of the input audio data. To this end,indexer220 may includespeaker segmentation logic221,speech recognition logic222,speaker clustering logic223,speaker identification logic224,name spotting logic225,topic classification logic226, andstory segmentation logic227.
• [0030]Speaker segmentation logic221 detects changes in speakers in a manner consistent with aspects of the present invention.Speaker segmentation logic221 is described in more detail below.
• [0031]Speech recognition logic222 may use statistical models, such as acoustic models and language models, to process input audio data. The language models may include n-gram language models, where the probability of each word is a function of the previous word (for a bi-gram language model) and the previous two words (for a tri-gram language model). Typically, the higher the order of the language model, the higher the recognition accuracy at the cost of slower recognition speeds. The language models may be trained on data that is manually and accurately transcribed by a human.
• [0032]Speaker clustering logic223 may identify all of the segments from the same speaker in a single document (i.e., a body of media that is contiguous in time (from beginning to end or from time A to time B)) and group them into speaker clusters.Speaker clustering logic223 may then assign each of the speaker clusters a unique label.Speaker identification logic224 may identify the speaker in each speaker cluster by name or gender.
• Name spotting[0033]logic225 may locate the names of people, places, and organizations in the transcription. Name spottinglogic225 may extract the names and store them in a database.Topic classification logic226 may assign topics to the transcription. Each of the words in the transcription may contribute differently to each of the topics assigned to the transcription.Topic classification logic226 may generate a rank-ordered list of all possible topics and corresponding scores for the transcription.
• [0034]Story segmentation logic227 may change the continuous stream of words in the transcription into document-like units with coherent sets of topic labels and other document features. This information may constitute metadata corresponding to the input audio data.Story segmentation logic227 may output the metadata in the form of documents tomemory system230, where a document corresponds to a body of media that is contiguous in time (from beginning to end or from time A to time B).
• In one implementation, logic[0035]222-227 may be implemented in a manner similar to that described by John Makhoul et al., “Speech and Language Technologies for Audio Indexing and Retrieval,” Proceedings of the IEEE, Vol. 88, No. 8, August 2000, pp.1338-1353, which is incorporated herein by reference.
• [0036]Memory system230 may store documents fromindexer220.Memory system230 may include one ormore databases231.Database231 may include a conventional database, such as a relational database, that stores documents fromindexer220.Database231 may also store documents received fromclients250 viaserver240.Server240 may include logic that interacts withmemory system230 to store documents indatabase231, query orsearch database231, and retrieve documents fromdatabase231.
• [0037]Server240 may include a computer or another device that is capable of interacting withmemory system230 andclients250 vianetwork260.Server240 may receive queries fromclients250 and use the queries to retrieve relevant documents frommemory system230.Clients250 may include personal computers, laptops, personal digital assistants, or other types of devices that are capable of interacting withserver240 to retrieve documents frommemory system230.Clients250 may present information to users via a graphical user interface, such as a web browser window.
• Typically, in the operation of[0038]system200, audio streams are transcribed as rich transcriptions that include metadata that defines information, such as speaker identification and story segments, related to the audio streams.Indexer220 generates the rich transcriptions.Clients250, viaserver240, may then search and browse the rich transcriptions.
• FIG. 3 is a diagram illustrating an[0039]exemplary computing device300 that may correspond toserver240 orclients250.Computing device300 may includebus310,processor320,main memory330, read only memory (ROM)340,storage device350,input device360,output device370, andcommunication interface380.Bus310 permits communication among the components ofcomputing device300.
• [0040]Processor320 may include any type of conventional processor or microprocessor that interprets and executes instructions.Main memory330 may include a random access memory (RAM) or another type of dynamic storage device that stores information and instructions for execution byprocessor320.ROM340 may include a conventional ROM device or another type of static storage device that stores static information and instructions for use byprocessor320.Storage device350 may include a magnetic and/or optical recording medium and its corresponding drive.
• [0041]Input device360 may include one or more conventional mechanisms that permit an operator to input information tocomputing device300, such as a keyboard, a mouse, a pen, a number pad, a microphone and/or biometric mechanisms, etc.Output device370 may include one or more conventional mechanisms that output information to the operator, including a display, a printer, speakers, etc.Communication interface380 may include any transceiver-like mechanism that enablescomputing device300 to communicate with other devices and/or systems. For example,communication interface380 may include mechanisms for communicating with another device or system via a network, such asnetwork260.
Exemplary Processing
• As previously mentioned,[0042]speaker segmentation logic221 locates positions in the input audio stream at which there are speaker changes. FIG. 4 is diagram illustrating functional components of speakerchange detection logic221. As shown, speakerchange detection logic221 includes inputstream segmentation component401, phoneclassification decode component402, and speaker change detection (SCD)component403.
• Input[0043]stream segmentation component401 breaks the input stream into a continuous stream of audio intervals. The audio intervals may be of a predetermined length, such as 30 seconds. FIG. 5 illustrates an audio stream broken into 30 second intervals501-504. Two successive intervals may include overlapping portions.Intervals502 and503, for example, may include overlappingportion505. Overlappingportion505 ensures that speakers changes at the boundaries of two intervals can still be detected.Portion505 is the end ofaudio interval502 and the beginning ofaudio interval503. Inputstream segmentation component401 passes successive intervals, such as intervals501-504, to phoneclassification decode component402.
• Returning to FIG. 4, phone[0044]classification decode component402 locates phones in each of its input audio intervals, such as intervals501-504. A phone is the smallest acoustic event that distinguishes one word from another. The number of phones used to represent a particular language may vary depending on the particular phone model that is employed. In the phone system described in Kubala et al., “The 1997 Byblos System Applied to Broadcast News Transcription,”Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop, Landsdowne, Va., February 1998, 45 context independent phone models were trained for both genders. Another silence phone model was trained for silence, giving a total of 91 phone models. The 91 phone models were then used to decode speech, resulting in an output sequence of phones with gender and silence labels.
• Phone[0045]classification decode component402, instead of attempting to decode audio intervals501-504 using a complete phone set, classifies audio intervals501-504 using a simplified phone decode set. More particularly, phoneclassification decode component402 may classify phones into three phone classes: vowels and nasals, fricatives or sibilants, and obstruents. The phones within each class have similar acoustic characteristics. Vowels and nasals are similar in that they both have pitch and high energy.
• Phone[0046]classification decode component402 may use four additional phones directed to non-speech events: music, laughter, breath and lip-smack, and silence. This gives a total of eight phone classes, which are illustrated in Table I, below.

  TABLE I
  	CONVENTIONAL PHONES
  CLASS	INCLUDED IN CLASS
  Vowels and Nasals	AX, IX, AH, EH, IH, OH, UH, EY, IY, AY, OY,
  	AW, OW, UW, AO, AA, AE, EI, ER, AXR, M,
  	N, NX, L, R, W, Y
  Fricatives	V, F, HH, TH, DH, Z, ZH, S, SH
  Obstruents	B, D, G, P, T, K, DX, JH, CH
  Music
  Laughter
  Breath and lip-smack
  Silence

• As shown in Table I, the phone class “vowels and nasals” includes a number of conventional phones. Similarly, the phone classes “fricatives” and “obstruents” may also include a number of conventional phones. Phone[0047]classification decode component402 does not, however, need to distinguish between the individual phones shown in any particular phone class. Instead, phoneclassification decode component402 simply classifies incoming audio as a sequence of the phone classes shown in Table I. This allows phoneclassification decode component402 to be trained on a reduced number of events (seven), thus requiring a simpler, and a more computationally efficient, phone decode model.
• Phone[0048]classification decode component402 may use a 5-state Hidden Markov Model (HMM) to model each of the seven phone classes. One codebook with 64 diagonal Gaussian Mixture Models (GMM) is shared by the 5 states from the same phone class and for each state a Gaussian mixture weight is trained. One suitable implementation of phoneclass decode component402 is discussed in Daben Liu et al., “Fast Speaker Change Detection for Broadcast News Transcription and Indexing,”Proceedings of Eurospeech99, Budapest, Hungary, September 1999, pp.1031-1034, which is incorporated herein by reference.
• FIG. 6 is a diagram illustrating exemplary phone classes[0049]601-605 decoded for an audio interval, such asinterval501, by phoneclassification decode component402. A typical 30 second speech interval may include approximately 300 phone classes. Between each phone class is a phone boundary610-612.
• Returning to FIG. 4,[0050]SCD component403 receives the phone-decoded audio intervals from phoneclassification decode component402.SCD component403 may then analyze each boundary610-612 to determine if the boundary corresponds to a speaker change. More specifically,SCD component403 compares the audio signal before the boundary and after the boundary within the audio interval. For example, when analyzingboundary611 ininterval501,SCD component403 may compare the audio signal corresponding tophone classes601 and602 with the audio signal corresponding to classes603-605.
• FIG. 7 is a flow chart illustrating the operation of[0051]SCD component403 in additional detail when detecting speaker change boundaries in intervals501-504 (FIG. 5).
• To begin,[0052]SCD component403 may set the starting position at one of the phone class boundaries of an interval to be analyzed (Act701). The selected boundary may not be the first boundary in the interval as there may not be sufficient audio data between the start of the interval and the first boundary for a valid comparison of acoustic features. Accordingly, the first selected boundary may be, for example, half-way into overlappingportion505.SCD component403 may similarly process the previous interval up to the point half-way into the same overlapping portion. In this manner, the complete audio stream is processed.
• [0053]SCD component403 may next calculate cepstral vectors for the regions before and after the selected boundary of the active interval (Act702). The calculation of cepstral vectors for samples of audio data is well known in the art and will not be described in detail herein. The two cepstral vectors are compared (Act703). In one implementation consistent with the present invention, the two cepstral vectors are compared using the generalized likelihood ratio test. Assume that the cepstral vector to the left of the selected boundary is vector x and the cepstral vector to the right of the selected boundary is vector y. The generalized likelihood ratio test may be written as:$\lambda =\frac{L\left(z;u_z,\sum _z\right)}{L\left(x;u_x,\sum _x\right)L\left(y;u_y,\sum _y\right)},$

  
• where L(v;u[0054]_v,Σ_v) is the maximum likelihood of v and z is the union of x and y.
• If λ is above a predetermined threshold, the two vectors are considered to be similar to one another, and are assumed to be from the same speaker ([0055]Acts704 and705). Otherwise, the two vectors are dissimilar to one another, and the boundary point corresponding to the two vectors is defined as a speaker change boundary (Acts704 and706). The predetermined threshold may be determined empirically.
• If there are any further boundaries in the interval,[0056]SCD component403 repeats Acts702-706 for the next boundary (Acts707 and708). If there are no further boundaries, the interval has been completely processed.
• In alternate implementations,[0057]SCD component403 may apply different threshold levels for boundaries that surround the non-speech phones. Thus, in between speech phones,SCD component403 may use a threshold that is less likely to give a speaker boundary change indication than a boundary between non-speech phones. Further, it is often the case that λ tends to be larger when calculating for larger data sets. Accordingly, a bias factor based on the size of the data set may be added to λ to compensate for this property.
CONCLUSION
• Speaker segmentation logic, as described herein, detects changes in speakers at boundary points determined by the location of phones in speech. The phones are decoded for the speech using a reduced set of possible phones. Further, the speaker segmentation logic processes the audio as discrete intervals of audio in which boundary portions of the audio are set to overlap to ensure accurate boundary detection over the whole audio signal.[0058]
• The foregoing description of preferred embodiments of the invention provides illustration and description, but is not intended to be exhaustive or to limit the invention to the precise form disclosed. Modifications and variations are possible in light of the above teachings or may be acquired from practice of the invention. Moreover, while a series of acts have been presented with respect to FIG. 7, the order of the acts may be different in other implementations consistent with the present invention.[0059]
• Certain portions of the invention have been described as software that performs one or more functions. The software may more generally be implemented as any type of logic. This logic may include hardware, such as application specific integrated circuit or a field programmable gate array, software, or a combination of hardware and software.[0060]
• No element, act, or instruction used in the description of the present application should be construed as critical or essential to the invention unless explicitly described as such. Also, as used herein, the article “a” is intended to include one or more items. Where only one item is intended, the term “one” or similar language is used.[0061]
• The scope of the invention is defined by the claims and their equivalents.[0062]

Claims (31)

What is claimed:

1. A method for detecting speaker changes in an input audio stream comprising:

segmenting the input audio stream into predetermined length intervals;

decoding the intervals to produce a set of phones corresponding to each of the intervals;

generating a similarity measurement based on a first portion of the audio stream within one of the intervals and prior to a boundary between adjacent phones and a second portion of the audio stream within the one of the intervals after the boundary; and

detecting speaker changes based on the similarity measurement.

2. The method ofclaim 1, wherein the predetermined length intervals are approximately thirty seconds in length.

3. The method ofclaim 1, wherein segmenting the input audio stream includes:

creating the predetermined length intervals such that portions of the intervals overlap one another.

4. The method ofclaim 1, wherein generating a similarity measurement includes:

calculating cepstral vectors for the audio stream prior to the boundary and the audio stream after the boundary, and

comparing the cepstral vectors.

5. The method ofclaim 4, wherein the cepstral vectors are compared using a generalized likelihood ratio test.

6. The method ofclaim 5, wherein a speaker change is detected when the generalized likelihood ratio test produces a value less than a preset threshold.

7. The method ofclaim 1, wherein the decoded set of phones is selected from a simplified corpus of phone classes.

8. The method ofclaim 7, wherein the simplified corpus of phone classes includes a phone class for vowels and nasals, a phone class for fricatives, and a phone class for obstruents.

9. The method ofclaim 8, wherein the simplified corpus of phone classes further includes a phone class for music, laughter, breath and lip-smack, and silence.

10. The method ofclaim 7, wherein the simplified corpus of phone classes includes approximately seven phone classes.

11. A device for detecting speaker changes in an audio signal, the device comprising:

a processor; and

a memory containing instructions that when executed by the processor cause the processor to:

segment the audio signal into predetermined length intervals,

decode the intervals to produce a set of phones corresponding to each of the intervals,

generate a similarity measurement based on a first portion of the audio signal prior to a boundary between phones in one of the sets of phones and a second portion of the audio signal after the boundary, and

detect speaker changes based on the similarity measurement.

12. The device ofclaim 11, wherein the predetermined length intervals are approximately thirty seconds in length.

13. The device ofclaim 11, wherein segmenting the audio signal includes:

creating the predetermined length intervals such that portions of the intervals overlap one another.

14. The device ofclaim 11, wherein the set of phones is selected from a simplified corpus of phone classes.

15. The device ofclaim 14, wherein the simplified corpus of phone classes includes a phone class for vowels and nasals, a phone class for fricatives, and a phone class for obstruents.

16. The device ofclaim 15, wherein the simplified corpus of phone classes further includes a phone class for music, laughter, breath and lip-smack, and silence.

17. The device ofclaim 11, wherein the simplified corpus of phone classes includes approximately seven phone classes.

18. A device for detecting speaker changes in an audio signal, the device comprising:

a segmentation component configured to segment the audio signal into predetermined length intervals;

a phone classification decode component configured to decode the intervals to produce a set of phone classes corresponding to each of the intervals, a number of possible phone classes being approximately seven; and

a speaker change detection component configured to detect locations of speaker changes in the audio signal based on a similarity value calculated over a first portion of the audio signal prior to a boundary between phone classes in one of the sets of phone classes and a second portion of the audio signal after the boundary in the one of the sets of phone classes.

19. The device ofclaim 18, wherein the predetermined length intervals are approximately thirty seconds in length.

20. The device ofclaim 18, wherein the segmentation component segments the predetermined length intervals such that portions of the intervals overlap one another.

21. The device ofclaim 18, wherein the phone classes include a phone class for vowels and nasals, a phone class for fricatives, and a phone class for obstruents.

22. The device ofclaim 21, wherein the phone classes further include a phone class for music, laughter, breath and lip-smack, and silence.

23. A system comprising:

an indexer configured to receive input audio data and generate a rich transcription from the audio data, the rich transcription including metadata that defines speaker changes in the audio data, the indexer including:

a segmentation component configured to divide the audio data into overlapping segments,

a speaker change detection component configured to detect locations of speaker changes in the audio data based on a similarity value calculated at locations in the segments that correspond to phone class boundaries;

a memory system for storing the rich transcription; and

a server configured to receive requests for documents and to respond to the requests by transmitting ones of the rich transcriptions that match the requests.

24. The system ofclaim 23, wherein the indexer further includes at least one of: a speaker clustering component, a speaker identification component, a name spotting component, and a topic classification component.

25. The system ofclaim 23, wherein the overlapping segments are segments of a predetermined length.

26. The system ofclaim 25, wherein the predetermined length is approximately thirty seconds.

27. The system ofclaim 23, wherein the phone classes include a phone class for vowels and nasals, a phone class for fricatives, and a phone class for obstruents.

28. The system ofclaim 27, wherein the phone classes additionally include a phone class for music, laughter, breath and lip-smack, and silence.

29. The system ofclaim 23, wherein the phone classes include approximately seven phone classes.

30. A device comprising:

means for segmenting the input audio stream into predetermined length intervals;

means for decoding the intervals to produce a set of phones corresponding to each of the intervals;

means for generating a similarity measurement based on audio within one of the intervals that is prior to a boundary between adjacent phones and based on audio within the one of the intervals that is after the boundary; and

means for detecting speaker changes based on the similarity measurement.

31. The device ofclaim 30, wherein the predetermined length intervals overlap one another.

Images