System and method for identifiying semantic intent from acoustic information

BACKGROUND OF THE INVENTION

0001The present invention deals with identifying semantic intent in acoustic information. More specifically, the present invention deals with grouping acoustic information (such as acoustic information from call logs) into clusters, each representing a category of semantic intent.

0002Automatic voice response systems have gained increasing popularity in enhancing human-machine interaction. Conventional automatic voice response systems allow a user to call the system using a telephone and then navigate through a voice-responsive menu in order to receive desired information, or to be routed to a desired destination. For instance, in some such systems, a user may call to review an account summary of the user's account with a particular business. In that case, the user may navigate through an account summary menu, using voice commands, to obtain an account balance, for example.

0003In another such system, the user may dial the general telephone number of a company and navigate through a voice-responsive menu to reach a particular individual at the company, or to reach a department, such as "technical service".

0004These types of systems have encountered a number of problems. In such systems, rules-based finite state or context free grammars (CFGs) are often used as a language model (LM) for simple, system-initiative dialog applications. This type of restricted strategy often leads to high recognition performance for in-grammar utterances, but completely fails when a user's response is not contained in the grammar.

0005There are at least two causes for such "out-of-grammar utterances". First, the syntactic structure of the utterance may not be parsed consistently by the CFG. For instance, a user's response of "twentieth of July" may cause failure in a grammar which is structured to include a rule [month] [day]. Second, the user's utterance may reflect a semantic intent which was not anticipated by the author of the grammar. For instance, in a corporate voice dialer application, the grammar for the response to the opening prompt "Good morning, who would you like to contact?" may be designed to expect the user to provide a name. However, the user may instead respond by identifying a department such as "human resources."

0006In sum, at the application design stage, it is difficult for an application developer to anticipate all the different ways in which a user may frame a request, which leads to the first problem. Similarly, it is difficult for an application developer to anticipate all the different semantic intents that the user may have, leading to the second problem.

0007Many attempts have been made to address the first problem (the difficulty in anticipating the different ways a user may frame a request) by building more robust language models. For example, hand-authored combinations of context free grammars (CFGs) with statistical language models has been attempted.

0008Prior attempts at solving the second problem (anticipating all the different semantic intents used by the user) typically require a large amount of transcribed and semantically annotated data from actual user calls. Of course, this tends to be extremely expensive to generate. For instance, in order to generate this type of semantically annotated data, the actual incoming calls must be recorded. Then, a human being must typically listen to all of these recordings in order to identify any semantic intents used by the caller, that were not yet expected or anticipated by the developer. However, a large company, which generates the call volumes necessary to obtain a useful quantity of data, may receive several thousand calls per day. Even if the human being only listens to the calls which failed in the interactive voice response unit (e.g., calls which ended in hang-ups) and if those calls only made up ten to twenty percent of the entire call volume, this would require the human to listen to hundreds of calls each day. This is extremely time consuming and expensive.

0009<nplcit id="ncit0001" npl-type="s"><text>Carlson, B. A. "Unsupervised topic clustering of switchboard speech messages" 1996 IEEE International Conference on Acoustics, Speech and Signal processing conference proceedings, IEEE New York, NY, USA, viol.1, 1996, pages 315-318</text></nplcit> discloses statistical techniques which can be used to automatically group speech data records based on the similarity of their content. The document further discloses a clustering algorithm relating to message representation, clustering and result analysis.

0010It is the object of the present invention to solve the afore-mentioned problems.

0011This object is solved by the subject matter of the independent claim(s).

0012Preferred embodiments are defined in the dependent claims.

0013In accordance with one embodiment of the present invention, unanticipated semantic intents are discovered in audio data in an unsupervised manner. For instance, the audio acoustics are clustered based on semantic intent and representative acoustics are chosen for each cluster. The human then need only listen to a small number of representative acoustics for each cluster (and possibly only one per cluster) in order to identify the unforeseen semantic intents.

0014The acoustics are subjected to speech recognition. The clustering is then performed on the speech recognition results, as opposed to the acoustics themselves. The developer may be able to identify unknown semantic intent by reviewing the speech recognition results.

0015In one embodiment, the developer need not even listen to any of the acoustics to identify unanticipated semantic intents. Instead, the new semantic intents can automatically be determined by tracking whether the acoustic clusters were recognized in the speech recognition process using the application grammar or a background grammar. If they were recognized using rules from the application grammar, then the semantic intent already exists in the application grammar and is not new. However, if they were recognized using a background grammar, then the semantic intent is not represented in the application grammar and is identified as a new, or unanticipated, semantic intent.

0016In accordance with an embodiment, the clusters are analyzed, automatically, and possible additional rules or revisions to the application grammars or language models in the human-machine interface (such as the AVR system) are automatically suggested.

BRIEF DESCRIPTION OF THE DRAWINGS

0017<ul id="ul0001" list-style="none" compact="compact"><li><figref idref="f0001">FIG. 1</figref> is a block diagram of one illustrative environment in which the present invention can be used.</li><li><figref idref="f0002">FIG. 2</figref> is a block diagram illustrating processing flow in a clustering system in accordance with one embodiment of the present invention.</li><li><figref idref="f0003">FIG. 3</figref> is a flow diagram illustrating the overall processing steps taken by the system shown in <figref idref="f0002">FIG. 2</figref>.</li><li><figref idref="f0004">FIG. 4</figref> is a flow diagram illustrating one illustrative method for language model-based clustering in more detail.</li><li><figref idref="f0005">FIG. 5</figref> is a flow diagram illustrating how clusters are initialized in more detail.</li></ul>

DETAILED DESCRIPTION OF ILLUSTRATIVE EMBODIMENTS

A. Computing System Overview

0018The present invention relates to identifying unforeseen or unanticipated, semantic intents in acoustic data. However, before discussing the present invention in greater detail, one illustrative environment in which the present invention can be used will be discussed.

0019<figref idref="f0001">FIG. 1</figref> illustrates an example of a suitable computing system environment 100 on which the invention may be implemented. The computing system environment 100 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 the computing environment 100 be interpreted as having any dependency or requirement relating to any one or combination of components illustrated in the exemplary operating environment 100.

0020The 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.

0021The 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.

0022With reference to <figref idref="f0001">FIG. 1</figref>, an exemplary system for implementing the invention includes a general purpose computing device in the form of a computer 110. Components of computer 110 may include, but are not limited to, a processing unit 120, a system memory 130, and a system bus 121 that couples various system components including the system memory to the processing unit 120. The system bus 121 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.

0023Computer 110 typically includes a variety of computer readable media. Computer readable media can be any available media that can be accessed by computer 110 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 by computer 100. Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier WAV 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.

0024The system memory 130 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 system 133 (BIOS), containing the basic routines that help to transfer information between elements within computer 110, such as during start-up, is typically stored in ROM 131. RAM 132 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by processing unit 120. By way o example, and not limitation, <figref idref="f0001">FIG. 1</figref> illustrates operating system 134, application programs 135, other program modules 136, and program data 137.

0025The computer 110 may also include other removable/non-removable volatile/nonvolatile computer storage media. By way of example only, <figref idref="f0001">FIG. 1</figref> illustrates a hard disk drive 141 that reads from or writes to non-removable, nonvolatile magnetic media, a magnetic disk drive 151 that reads from or writes to a removable, nonvolatile magnetic disk 152, and an optical disk drive 155 that reads from or writes to a removable, nonvolatile optical disk 156 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. The hard disk drive 141 is typically connected to the system bus 121 through a non-removable memory interface such as interface 140, and magnetic disk drive 151 and optical disk drive 155 are typically connected to the system bus 121 by a removable memory interface, such as interface 150.

0026The drives and their associated computer storage media discussed above and illustrated in <figref idref="f0001">FIG. 1</figref>, provide storage of computer readable instructions, data structures, program modules and other data for the computer 110. In <figref idref="f0001">FIG. 1</figref>, for example, hard disk drive 141 is illustrated as storing operating system 144, application programs 145, other program modules 146, and program data 147. Note that these components can either be the same as or different from operating system 134, application programs 135, other program modules 136, and program data 137. Operating system 144, application programs 145, other program modules 146, and program data 147 are given different numbers here to illustrate that, at a minimum, they are different copies.

0027A user may enter commands and information into the computer 110 through input devices such as a keyboard 162, a microphone 163, and a pointing device 161, 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 the processing unit 120 through a user input interface 160 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 urliversal serial bus (USB). A monitor 191 or other type of display device is also connected to the system bus 121 via an interface, such as a video interface 190. In addition to the monitor, computers may also include other peripheral output devices such as speakers 197 and printer 196, which may be connected through an output peripheral interface 190.

0028The computer 110 may operate in a networked environment using logical connections to one or more remote computers, such as a remote computer 180. The remote computer 180 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 the computer 110. The logical connections depicted in <figref idref="f0001">FIG. 1</figref> 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.

0029When used in a LAN networking environment, the computer 110 is connected to the LAN 171 through a network interface or adapter 170. When used in a WAN networking environment, the computer 110 typically includes a modem 172 or other means for establishing communications over the WAN 173, such as the Internet. The modem 172, which may be internal or external, may be connected to the system bus 121 via the user-input interface 160, or other appropriate mechanism. In a networked environment, program modules depicted relative to the computer 110, or portions thereof, may be stored in the remote memory storage device. By way of example, and not limitation, <figref idref="f0001">FIG. 1</figref> illustrates remote application programs 185 as residing on remote computer 180. 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.

B. Acoustic Processing System Overview

0030<figref idref="f0002">FIG. 2</figref> illustrates an acoustic processing system 200 that can be used to cluster acoustics by semantic intent and to optionally suggest updates to a grammar in order to accommodate unanticipated semantic intents. System 200 includes a clustering system 202 which, itself, includes language model based clustering system 204, ranking and filtering system 206, and optional grammar updating system 208. <figref idref="f0002">FIG. 2</figref> also shows that clustering system 202 is coupled to speech recognition system 210 and an automatic voice response (AVR) system (or application) 212.

0031AVR system (or application) 212 is illustratively a human-machine interface that receives voice commands from a human being and attempts to take action based on those commands. In one illustrative embodiment, the voice commands are received by telephone. AVR system also illustratively logs (or stores) the acoustic data representative of the received audio commands. In one specific embodiment, AVR system 212 is an automatic attendant system deployed at a company to receive and direct calls.

0032Speech recognition system 210 is illustratively a conventional speech recognition system, and illustratively uses acoustic models that are the same as those used in clustering system 202, described below. Speech recognition system 210 illustratively employs a large vocabulary such that it is a large, generalized vocabulary recognizer. Alternatively, speech recognition system 210 can include an in-domain (or context-specific) recognizer in conjunction with a large, generalized vocabulary recognizer.

0033Clustering system 202 clusters the stored acoustics, based on the speech recognition results. Each cluster is illustratively indicative of a semantic intent expressed by the acoustics in that cluster. System 202 can also, in one embodiment, suggest revisions to the application grammar in AVR system 212.

0034<figref idref="f0003">FIG. 3</figref> is a flow diagram illustrating the overall operation of system 200 shown in <figref idref="f0002">FIG. 2</figref>. First, the acoustic information from desired call logs in AVR system 212, is extracted. The call log information is represented by block 214 in <figref idref="f0002">FIG. 2</figref> and the extraction step is represented by block 216 in <figref idref="f0003">FIG. 3</figref>. The desired call log information to be extracted is represented by acoustic waveforms for which semantic intents are to be recovered. These call logs of interest can be identified in a wide variety of ways. For example, the call logs of interest can be the acoustic information corresponding to failed calls in AVR system 212. It can also be a response to a particular prompt from all calls that failed in AVR system 212. By failed, it is generally meant that the caller prematurely hung-up (e.g., hung-up prior to achieving a desired objective or reaching a desired dialog state in AVR system 212). For instance, in one specific example, the call log information of interest can be the acoustic waveforms provided in response to a prompt "Good morning, who would you like to contact?" for all calls that ended with the caller hanging up prior to completing a task or speaking with a person. Of course, the call log information of interest 212 can also be all acoustics recorded over a given period of time, or a different subset of those acoustics, as desired.

0035Once the call log information of interest has been extracted, it is provided to speech recognition system 210 where speech recognition is performed on the extracted acoustics. The speech recognition results are indicated by block 218 in <figref idref="f0002">FIG. 2</figref> and the performance of speech recognition on the call log information 214 is indicated by block 220 in <figref idref="f0003">FIG. 3</figref>.

0036Speech recognition results 218 can take one of a variety of different forms. For instance, results 218 can be the one-best hypothesis recognized by speech recognition system 210, the n-best hypotheses or a recognition lattice, all of which are known types of outputs from speech recognition systems. It is, of course, important that speech recognition system 210 cover words that are outside the application grammar used by AVR system 212. This is to ensure that most words in the new or unanticipated semantic intents expressed in the extracted call logs are covered and can be recognized by speech recognition system 210. However, it is not necessary that all words be within the grammar coverage of speech recognition system 210, nor is it necessary to have all waveforms correctly recognized. Word level recognition can be used in the present clustering system, even if they are inaccurate recognition results, so long as acoustic waveforms with similar semantics have consistent recognition results. For instance, as long as acoustic waveforms representing the phrase "good morning" are recognized consistently as "get morning" these results can be used by clustering system 202, even though they are incorrect. Speech recognition results 218 are provided to clustering system 202, and specifically to language model-based clustering system 204. The detailed operation of language model-based clustering system 204 is described later with respect to <figref idref="f0004">FIGS. 4</figref> and <figref idref="f0005">5</figref>. Suffice it to say, for now, that the acoustic call log information of interest 214 which was extracted from AVR system 212 is clustered based on the semantic intent represented by speech recognition results 218 corresponding to that acoustic information, using a language model clustering approach implemented by system 204. For instance, each acoustic waveform is represented by its recognition results 218. A cluster is modeled by a per-cluster generative language model, which is a generative model of word sequences. The probability of an acoustic waveform, given a cluster, is basically the probability of its recognition results given the cluster language model. This is described in greater detail below with respect to <figref idref="f0004">FIGS. 4</figref> and <figref idref="f0005">5</figref>.

0037The semantically based clusters 222 are output by system 204. The performance of language model-based clustering of acoustics based on speech recognition results 218 is indicated by block 224 in <figref idref="f0003">FIG. 3</figref>.

0038Clusters 222 are then ranked and filtered by system 206. The clustering performed by clustering system 204 may result in a significant number of clusters. Therefore, it may be important to select certain of those clusters for presentation to an application developer, in order to save time and resources. This involves ranking the clusters in order of importance, filtering out unimportant or "garbage" clusters and representing a cluster in a simple and relatively self-descriptive way.

0039In accordance with one embodiment of the present invention, clusters 222 are ranked based on their frequency (i.e., based on the number of instances of utterances contained in a cluster). This information indicates how frequently a semantic intent occurs in the dataset.

0040Once the clusters 222 are ranked based on frequency, they are filtered. A cluster with a high frequency may not necessarily be relevant. For instance, there may be a relatively high number of calls that consist only of silence, noise, or other incoherent speech. These "garbage" utterances tend to be recognized as some certain function words or word sequences such as "a", "oh", "the", for example. They are likely to be clustered together with a high cluster prior count. However, unlike utterances in a cluster with meaningful semantics, these garbage word sequences are seldom consistent with one another.

0041Therefore, in accordance with one embodiment of the present invention, a "consistency" measure is used to filter out garbage clusters. This metric can also be referred to as "compactness" as it is computed in an attempt to pick out those clusters with a large portion of constant instances, and to identify a "center" instance to represent the generative cluster. In one embodiment, a similarity measure is first defined between two utterances to be the number of word tokens they have in common, normalized by the total number of word tokens in both of their n-best decoding results. The "consistency" is then defined as the normalized, pair-wise similarity of all utterances in a cluster. The clusters with a consistency lower than a threshold value are considered "garbage" and are discarded. The threshold value can be empirically determined.

0042It will be recognized that there is a trade-off in setting the consistency threshold. If it set relatively high, then this enhances the likelihood that only relevant clusters will meet the consistency threshold, but the system may then discard some important or relevant clusters. If the threshold is set relatively low, then it is unlikely that the system will miss or filter out any relevant clusters, but it is more likely that it will include some garbage clusters.

0043Once ranking and filtering system 206 has ranked and filtered the clusters, it selects a central utterance to represent each remaining cluster. This utterance can be chosen to have a highest sum of similarities with all other utterances in the same cluster, or it can be chosen in other ways as well. This will likely turn out to be intuitively the most representative utterance in the cluster. The distance measure for "similarity" will illustratively be the same as that used to define consistency when filtering the clusters.

0044The selected clusters output by system 206 are represented by block 226 in <figref idref="f0002">FIG. 2</figref>, and the process of ranking and filtering clusters to obtain the selected clusters 208 is indicated by block 228 in <figref idref="f0003">FIG. 3</figref>.

0045In one illustrative embodiment, clustering system 202 is finished after this step and simply outputs the selected clusters 226 for developer review. This is indicated by block 230 in <figref idref="f0003">FIG. 3</figref>. In this embodiment (in which the selected clusters 226 are output), the present invention effectively helps an application developer to reduce the number of clusters which the developer is required to review, but the application developer still needs to decide, by inspecting the representative utterance(s) for each selected cluster 226 (or by listening to the corresponding acoustics if the speech recognition results are erroneous) whether the cluster has an unanticipated semantic intent or whether it has one that already exists in the application grammar used by AVR system 212. Again, in this embodiment, once one of the selected clusters 226 is decided to have a new semantic intent, the application developer illustratively generates any necessary corrections to the word sequences in the cluster (in case the speech recognition results were erroneous as discussed above) and can use the cluster to learn or generate a new grammar rule based on the corrected word sequences. The new grammar rule can be generated automatically or manually. The updated grammar or rule can then be used to update the grammar used by AVR system 212 in order to accommodate the previously unanticipated semantic intent.

0046However, it will also be appreciated that the present clustering system 202 can include optional grammar updating system 208 which automatically generates a new grammar rule or updates the application grammar based on the selected clusters 226. One suggested update to the application grammar can simply be the language model generated for the cluster. The top rule for the application grammar will then be given a weight which may illustratively be one minus the sum of the weights of all other newly discovered rules representing all other semantic intents. The new rule or grammar will thus be integrated into the existing application grammar and can be used by AVR system 212. The updated application grammar is indicated by block 232 in <figref idref="f0002">FIG. 2</figref>, and the process of automatically updating the grammar used by the application is indicated by block 234 in <figref idref="f0003">FIG. 3</figref>.

0047A number of modifications can also be made to the embodiments described herein in order to assist the developer. For instance, where the selected clusters are output to the developer for review, the developer needs to decide which clusters are already represented by the application grammar and which are new (or were unantioipated). In order to do this, speech recognition system 210 may employ not only the large vocabulary recognizer, but may also employ the application grammar used by AVR system 212. In that embodiment, if the speech recognition results 218 were generated using the large vocabulary grammar (or background grammar), but not the application grammar, they can be tagged as such and therefore easily identified as representing a new semantic intent (one not previously anticipated by the grammar used by AVR system 212). However, if the speech recognition results 218 were generated by speech recognition system 210 using the application grammar used by AVR system 212, then they can be tagged as such and easily identified as representing a semantic intent that is already covered by the application grammar.

0048If the results are tagged in this way, then the clusters can be identified as representing unanticipated semantic intent or previously covered semantic intent by simply counting the number of utterances in each cluster that have speech recognition results that were generated using the application grammar and those generated using the background grammar. If most of the utterances in a given cluster were generated using the background grammar, the developer may wish to determine that the cluster represents an unanticipated semantic intent. Alternatively, if most utterances corresponding to the cluster were generated using the application grammar, the developer may wish to determine that the semantic intent represented by that cluster is already covered by the application grammar. Of course, different schemes or thresholds can be used, as desired, in order to determine whether the cluster represents a new or existing semantic intent.

0049For instance, speech recognition results generated from the different grammars are not likely to be clustered together, since they likely do not have many lexicon items in common. Therefore, each cluster will likely have a pronounced majority of recognition results generated from one grammar, but not both. Therefore, the tag of the representative utterance may be sufficient to indicate whether the cluster represents known or unanticipated semantic intent.

0050Alternatively, instead of only clustering calls that ended in failure, the acoustic information for all calls to AVR system 212 can be used in accordance with the present invention, even if the calls succeeded. The acoustics corresponding to calls that failed can easily be tagged, as can the acoustic scores corresponding to calls that succeeded. The utterances represented by the acoustics tagged as corresponding to calls that succeeded can be assumed to contain semantic intent that is already covered by the application grammar. Those tagged as corresponding to calls that failed can be assumed to contain semantic intent that is not anticipated by the application grammar. It will of course be readily appreciated that this does not require the application grammar to be employed by the speech recognition system 210, but it still allows the grammar updating system 208 to automatically determine whether a cluster represents unanticipated semantic intent or semantic intent that is already know by system 212.

C. Clustering in More Detail

0051<figref idref="f0004">FIG. 4</figref> is a flow diagram which illustrates one illustrative embodiment of the particular language model-based clustering approach used by clustering system 204, in greater detail. In one illustrative embodiment, a generative Markov model is used, where the acoustic feature sequence <i>x</i> of an utterance is generated from a word sequence w according to an acoustic model <i>p(x</i>|<i>w)</i>, and a word sequence w is generated from a semantic intent (or cluster) <i>c</i> based on a per-cluster n-gram language model <i>p(w</i>|<i>c)</i>. The complete likelihood of <i>x,w</i> and <i>c</i> then becomes: <maths id="math0001" num="Eq. 1"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>1</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>x</mi><mi>w</mi><mi>c</mi></mfenced><mo>=</mo><mi>p</mi><mfenced><mi>x</mi><mrow><mo>|</mo><mi>w</mi></mrow></mfenced><mo>⁢</mo><mi>p</mi><mfenced><mi>w</mi><mrow><mo>|</mo><mi>c</mi></mrow></mfenced><mo>⁢</mo><mi>p</mi><mfenced><mi>c</mi></mfenced><mo>,</mo></mtd></mtr></mtable></math><img file="EP1669980B1_D0001.tif" /></maths>

0052The present system illustratively trains models corresponding to semantic clusters so as to maximize the likelihood <i>p(x)</i>. In one illustrative embodiment, a fixed acoustic model <i>p(x</i>|<i>w)</i> is used in clustering. This model is trained offline on a large set of telephony speech. Per-cluster uni-grams can be used to model <i>p(w</i>|<i>c)</i>, where the sentence end probability is set to be equal among all clusters.

0053As previously mentioned, semantic intents are often expressed by very short utterances in telephony applications. Therefore, uni-grams can be chosen because it is believed that in such applications, a uni-gram language model corresponding to a semantic cluster has a perplexity that is not much higher than a bi-gram (or tri-gram) language model, but has a much lower computational complexity. Therefore, training in accordance with the present invention involves estimating the alphabet of the cluster <i>c</i>, the prior probability for semantic clusters <i>p(c)</i>, and the language models <i>p(w</i>|<i>c)</i>.

1. Model Initialization

0054Before discussing estimation of the language models in more detail, it should first be noted that model initialization can be important in unsupervised clustering. Therefore, the first step is to initialize models corresponding to the clusters. This is indicated by block 300 shown in <figref idref="f0004">FIG. 4</figref>. The process of initializing the clusters is indicated in greater detail by the flow diagram shown in <figref idref="f0005">FIG. 5</figref>, which will now be discussed, before the discussion proceeds with respect to the remaining blocks in <figref idref="f0004">FIG. 4</figref>.

0055In order to initialize the clusters, the language model based clustering system 204 first enumerates all vocabulary items in the speech recognition results 218. This is indicated by block 302 in <figref idref="f0005">FIG. 5</figref>. For instance, assume that the speech recognition results included the utterances "operator", "ACME operator", and "the operator". A cluster is thus initialized corresponding to each of the words "operator", "ACME", and "the", since these are the lexical items (or words) contained in the speech recognition results 218. Therefore, the number of clusters created is the same as the number of vocabulary items that have a count no less than a floor count (in the present example, the floor count is one) in the speech recognition results 218, each cluster corresponding to one of the vocabulary items (or words) in the results. Creating a cluster for each enumerated vocabulary item is indicated by block 304 in <figref idref="f0005">FIG. 5</figref>.

0056The speech recognition results that contain these lexical items are then assigned to each of the clusters. For instance, since the speech recognition result "operator" contains the word "operator", that utterance will be assigned only to the cluster created for the word "operator". The utterance "ACME operator", on the other hand, will be assigned to both the cluster created for the word "operator" and the cluster created for the word "AMCE", since it contains both words. Similarly, the utterance "the operator" will be assigned both to the cluster created for the word "the" and the cluster created for the word "operator".

0057The prior probability for each cluster <i>p(c)</i> corresponding to a word <i>v</i> is set to the normalized number of utterances containing v in that cluster. This is indicated by block 306 in <figref idref="f0005">FIG. 5</figref>. The instances of utterances are then assigned to the clusters. This is indicated by block 308 in <figref idref="f0005">FIG. 5</figref>. An n-gram language model <i>p(w</i>|<i>c)</i> for each cluster is then trained based on the word sequences w_i* containing the word v (i.e., based on the instances assigned to that cluster). This is indicated by block 310 in <figref idref="f0005">FIG. 5</figref>.

2. Refining the Clusters

0058Once the clusters and language models are initialized as described with respect to <figref idref="f0005">FIG. 5</figref>, processing continues in <figref idref="f0004">FIG. 4</figref> where the clusters are refined and those having insufficient counts are removed. This is indicated by block 312 in <figref idref="f0004">FIG. 4</figref>.

0059Refining the clusters is performed by maximizing the likelihood of an acoustic dataset {<i>x_i</i>}<i>^M</i> consisting of <i>M</i> waveforms <i>x_i</i>. Since <i>w</i> and <i>c</i> are hidden, the EM algorithm can be applied to train the models. This can be done by reassigning each utterance <i>x_i</i> to a cluster by finding the posterior probability: <maths id="math0002" num="Eq. 2"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>2</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mfrac><mstyle displaystyle="true"><munder><mo>∑</mo><mi>w</mi></munder><mi>p</mi><mfenced><mi>c</mi></mfenced><mi>p</mi><mfenced><mi>w</mi><mrow><mo>|</mo><mi>c</mi></mrow></mfenced><mi>p</mi><mfenced><msub><mi>x</mi><mi>i</mi></msub><mrow><mo>|</mo><mi>w</mi></mrow></mfenced></mstyle><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>c</mi><mo>′</mo></msup></munder><mstyle displaystyle="true"><munder><mo>∑</mo><mi>w</mi></munder><mi>p</mi><mfenced><msup><mi>c</mi><mo>′</mo></msup></mfenced><mi>p</mi><mfenced><mi>w</mi><mrow><mo>|</mo><msup><mi>c</mi><mo>′</mo></msup></mrow></mfenced><mi>p</mi><mfenced><msub><mi>x</mi><mi>i</mi></msub><mrow><mo>|</mo><mi>w</mi></mrow></mfenced></mstyle></mstyle></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0002.tif" /></maths> where c a is specific cluster and c' is a variable representing cluster <i>i</i> such that the sum over c' means summing over all clusters.

0060Since the sum over the word sequence w at each iteration is impractical, offline recognition can be employed with a background language model (as opposed to recognition at each iteration using Σ<i>p</i>(<i>c</i>)<i>p</i>(<i>w</i>|<i>c</i>)). Then, <maths id="math0003" num="Eq. 3"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>3</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mfrac><mrow><mi>p</mi><mfenced><mi>c</mi></mfenced><mo>⁢</mo><mi>p</mi><mfenced><msubsup><mi>w</mi><mi>i</mi><mo>∗</mo></msubsup><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mrow><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>c</mi><mo>′</mo></msup></munder><mi>p</mi><mfenced><msup><mi>c</mi><mo>′</mo></msup></mfenced><mi>p</mi><mfenced><msubsup><mi>w</mi><mi>i</mi><mo>∗</mo></msubsup><mrow><mo>|</mo><msup><mi>c</mi><mo>′</mo></msup></mrow></mfenced></mstyle></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0003.tif" /></maths> where <i>w_i*</i> is the recognition result for <i>x_i</i>.

0061An N-best list or lattice can be used where the N-best list for <i>x_i</i> is <i>w_i</i>,<i>₁</i>,..., <i>w_i</i>,<i>_j</i>,... <i>w_i</i>,<i>_N</i>, along with posterior probabilities <i>p(w_ij</i>|<i>x_i)</i> where: <maths id="math0004" num="Eq. 4"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>4</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mfrac><mrow><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mfenced><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>x</mi><mi>i</mi></msub><mrow><mo>|</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mrow></mfenced></mrow><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><msup><mi>j</mi><mo>′</mo></msup><mo>=</mo><mn>1</mn></mrow><mi>N</mi></munderover><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><msup><mi>j</mi><mo>′</mo></msup></mrow></msub></mfenced><mi>p</mi><mfenced><msub><mi>x</mi><mi>i</mi></msub><mrow><mo>|</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><msup><mi>j</mi><mo>′</mo></msup></mrow></msub></mrow></mfenced></mstyle></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0004.tif" /></maths> where <i>p(w_ij)</i> represents the background language model and <i>p(x_i</i>|<i>w_ij)</i> represents the acoustic model. Then, <maths id="math0005" num="Eq. 5"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>5</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>j</mi></munder><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mrow></mfenced></mstyle><mo>=</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>j</mi></munder><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mfrac><mrow><mi>p</mi><mfenced><mi>c</mi></mfenced><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mrow><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>c</mi><mo>′</mo></msup></munder><mi>p</mi><mfenced><msup><mi>c</mi><mo>′</mo></msup></mfenced><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msup><mi>c</mi><mo>′</mo></msup></mrow></mfenced></mstyle></mfrac></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0005.tif" /></maths>

0062For the embodiment in which a lattice is used, the sum over <i>j</i> can be implemented efficiently using a forward/backward algorithm.

0063We can also find: <maths id="math0006" num="Eq. 6"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>6</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mfrac><mrow><mi>p</mi><mfenced><mi>c</mi></mfenced><mo>⁢</mo><mi>p</mi><mfenced><mi>w</mi><mrow><mo>|</mo><mi>c</mi></mrow></mfenced><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>x</mi><mi>i</mi></msub><mrow><mo>|</mo><mi>w</mi></mrow></mfenced></mrow><mstyle displaystyle="true"><munder><mo>∑</mo><mrow><msup><mi>c</mi><mo>′</mo></msup><mo>,</mo><msup><mi>w</mi><mo>′</mo></msup></mrow></munder><mi>p</mi><mfenced><msup><mi>c</mi><mo>′</mo></msup></mfenced><mi>p</mi><mfenced><msup><mi>w</mi><mo>′</mo></msup><mrow><mo>|</mo><msup><mi>c</mi><mo>′</mo></msup></mrow></mfenced><mi>p</mi><mfenced><msub><mi>x</mi><mi>i</mi></msub><mrow><mo>|</mo><msup><mi>w</mi><mo>′</mo></msup></mrow></mfenced></mstyle></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0006.tif" /></maths> which can be approximated using recognition as: <maths id="math0007" num="Eq. 7"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>7</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><mi>x</mi><mo>⁢</mo><mi>i</mi></mrow></mfenced></mtd><mtd><mi>if</mi><mi> </mi><mi>w</mi><mo>=</mo><msubsup><mi>w</mi><mi>i</mi><mo>∗</mo></msubsup></mtd></mtr><mtr><mtd><mn>0</mn></mtd><mtd><mi>if</mi><mi> </mi><mi>w</mi><mo>≠</mo><msubsup><mi>w</mi><mi>i</mi><mo>∗</mo></msubsup></mtd></mtr></mtable></mrow></mtd></mtr></mtable></math><img file="EP1669980B1_D0007.tif" /></maths> or using an N-best list or lattices as: <maths id="math0008" num="Eq. 8"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>8</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mrow></mfenced><mo>=</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mfrac><mrow><mi>p</mi><mfenced><mi>w</mi></mfenced><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mrow><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>c</mi><mo>′</mo></msup></munder><mi>p</mi><mfenced><msup><mi>c</mi><mo>′</mo></msup></mfenced><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msup><mi>c</mi><mo>′</mo></msup></mrow></mfenced></mstyle></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0008.tif" /></maths> when <i>w</i> is the N-best list, and otherwise: <maths id="math0009" num="Eq. 9"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>9</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mn>0</mn></mtd></mtr></mtable></math><img file="EP1669980B1_D0009.tif" /></maths>

0064We now compute the following counts where #<i>_u</i>(<i>w</i>) is defined as the number of times that the word token occurs in the utterance w: <maths id="math0010" num="Eq. 10"><math display="block"><mtable columnalign="left"><mtr><mtd><msub><mi>#</mi><mi>c</mi></msub></mtd><mtd><mo>=</mo><msub><mi mathvariant="normal">Ψ</mi><mi>c</mi></msub><mo>=</mo><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>M</mi></munderover><mi>p</mi><mfenced><mi>c</mi><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mfenced></mtd></mtr><mtr><mtd><msub><mi>#</mi><mrow><mi>c</mi><mo>,</mo><mi>u</mi></mrow></msub></mtd><mtd><mo>=</mo><msub><mi>φ</mi><mrow><mi>c</mi><mo>,</mo><mi>u</mi></mrow></msub><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>M</mi></munderover><mstyle displaystyle="true"><mstyle displaystyle="false"><mstyle displaystyle="true"><munder><mo>∑</mo><mi>w</mi></munder></mstyle><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><msub><mo>#</mo><mi>u</mi></msub><mfenced><mi>w</mi></mfenced></mstyle></mstyle></mstyle></mtd></mtr><mtr><mtd><mspace width="1em" /></mtd><mtd><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>M</mi></munderover><mstyle displaystyle="true"><mstyle displaystyle="false"><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><msub><mo>#</mo><mi>u</mi></msub><mfenced><msubsup><mi>w</mi><mi>i</mi><mo>*</mo></msubsup></mfenced></mstyle></mstyle></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0010.tif" /></maths> or <maths id="math0011" num="Eq. 11"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>11</mn></mtd></mtr><mtr><mtd><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>M</mi></munderover><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>N</mi></munderover><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mrow></mfenced><mo>⁢</mo><msub><mo>#</mo><mi>u</mi></msub><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mfenced></mstyle></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0011.tif" /></maths>

0065Computing these expected counts ψ<i>_c</i> and φ<i>_c,u</i> corresponds to the E-step of the EM algorithm which provides sufficient statistics for the likelihood maximization. The <i>M</i> step thus simply includes normalizing φ<i>_c</i> to give the cluster prior probabilities <i>p(c)</i>, and normalizing φ<i>_c,u</i> to give the class-conditional uni-gram probabilities.

0066In other words, <i>p(c), p(w</i>|<i>c)</i> with <i>p'(c)</i>, <i>p'(w</i>|<i>c)</i> as follows: <maths id="math0012" num="Eq. 12"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>12</mn></mtd></mtr><mtr><mtd><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>c</mi></mfenced><mo>=</mo><mfrac><msub><mi>ψ</mi><mi>c</mi></msub><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>c</mi><mo>′</mo></msup></munder><msub><mi>ψ</mi><msup><mi>c</mi><mo>′</mo></msup></msub></mstyle></mfrac><mo>=</mo><mfrac><msub><mi>ψ</mi><mi>c</mi></msub><mi>M</mi></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0012.tif" /></maths>

0067Since <i>p(w</i>|<i>c)</i> is a uni-gram: <maths id="math0013" num="Eq. 13"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>13</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>w</mi><mrow><mo>|</mo><mi>c</mi></mrow></mfenced><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∏</mo><mrow><mi>k</mi><mo>=</mo><mn>1</mn></mrow><mrow><mi>ℓ</mi><mfenced><mi>w</mi></mfenced></mrow></munderover><msub><mi>p</mi><mi>c</mi></msub><mfenced><mi>w</mi><mfenced><mi>k</mi></mfenced></mfenced></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0013.tif" /></maths> where <i>w(k)</i> is the <i>k^th</i> word in sentence <i>w</i>, ℓ(<i>w</i>) is the length of sentence <i>w</i>, and <i>p_c(v)</i> is the uni-gram probability of word <i>v</i> in class <i>c</i>. Finally: <maths id="math0014" num="Eq. 14"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>14</mn></mtd></mtr><mtr><mtd><msub><msup><mi>p</mi><mo>′</mo></msup><mi>c</mi></msub><mfenced><mi>v</mi></mfenced><mo>=</mo><mfrac><msub><mi>ψ</mi><mrow><mi>c</mi><mo>,</mo><mi>v</mi></mrow></msub><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>v</mi><mo>′</mo></msup></munder><msub><mi>ψ</mi><mrow><mi>c</mi><mo>,</mo><msup><mi>v</mi><mo>′</mo></msup></mrow></msub></mstyle></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0014.tif" /></maths>

0068In actual implementation, the language model used in recognition (i.e., <i>p(w)</i> in computing <i>p(w</i>|<i>x)=p(x</i>|<i>w)p(w))</i> is decoupled from the per-cluster uni-gram language models (i.e., <i>p(w</i>|<i>c))</i>. In one specific embodiment, a task-independent large vocabulary background language model is used to compute <i>p(w)</i>. This has the advantage that with the language model <i>p(w)</i> and the acoustic model p(x|w) fixed, the recognition is performed offline, only once. The obtained word sequence hypotheses and their acoustic scores are used directly in training the clusters.

0069In addition, in one specific embodiment, for computational efficiency, the word sequence hypothesis is restricted to a lattice, or N-best list, with <i>p(w</i>|<i>x_i)</i> renormalized accordingly. In one aggressive embodiment, wherein an N-best list of length 1 is used, 1-best word sequence is obtained.

3 . Model Refining With Viterbi Training

0070In addition, it should be noted that Viterbi training can be used instead of EM training to optimize cluster parameters. In other words, <i>p(c</i>|<i>w)</i> is renormalized to 0 or 1, depending on whether c is the best hypothesis giver <i>w</i>.

0071More specifically, the following can be used: <maths id="math0015" num="Eq. 15"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>15</mn></mtd></mtr><mtr><mtd><mover accent="true"><mi>p</mi><mo></mo></mover><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mn>1</mn></mtd><mtd><mi>if</mi><mi> </mi><mi>c</mi><mi> </mi><mi>maximizes</mi><mi> </mi><mi>p</mi><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced></mtd></mtr><mtr><mtd><mn>0</mn></mtd><mtd><mi>otherwise</mi></mtd></mtr></mtable></mrow></mtd></mtr></mtable></math><img file="EP1669980B1_D0015.tif" /></maths> instead of <i>p</i>(<i>c</i>|<i>x_i</i>) to do "hard" or Viterbi class assignment. This can be more efficient.

0072For the 1-best case we then have: <maths id="math0016" num="Eq. 16"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>16</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mi>p</mi></mtd><mtd><mn>1</mn><mi> </mi><mi>if</mi><mi> </mi><mi>c</mi><mi> </mi><mi>is the best class and</mi><mi> </mi><mi>w</mi><mo>=</mo><msup><mi>w</mi><mo>∗</mo></msup></mtd></mtr><mtr><mtd><mn>0</mn></mtd><mtd><mi>otherwise</mi></mtd></mtr></mtable></mrow><mo>.</mo></mtd></mtr></mtable></math><img file="EP1669980B1_D0016.tif" /></maths>

0073For the embodiment in which an N-best list or lattice is used, there are two options. The first option is to choose one class as an overall choice for all hypotheses. To do this, let <i>p̃</i>(<i>c</i>|<i>x_i</i>) place a probability of one on maximizer of <i>p</i>(<i>c</i>|<i>x_i</i>) given in Equation 5 above. Then: <maths id="math0017" num="Eq. 17"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>17</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mover accent="true"><mi>p</mi><mo></mo></mover><mfenced><mi>c</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mfrac><mrow><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mrow><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><msup><mi>j</mi><mo>′</mo></msup><mo>=</mo><mn>1</mn></mrow><mi>N</mi></munderover><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><msup><mi>j</mi><mo>′</mo></msup></mrow></msub><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mstyle></mfrac></mtd><mtd><mi>if</mi><mi> </mi><mi>w</mi><mo>=</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub></mtd></mtr><mtr><mtd><mn>0</mn></mtd><mtd><mi>otherwise</mi></mtd></mtr></mtable></mrow></mtd></mtr></mtable></math><img file="EP1669980B1_D0017.tif" /></maths>

0074The second option is to choose a class per hypothesis. This can be done by letting: <maths id="math0018" num="Eq. 18"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>18</mn></mtd></mtr><mtr><mtd><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced></mtd><mtd><mi>if</mi><mi> </mi><mi>w</mi><mo>=</mo><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mi> </mi><mi>and</mi><mi> </mi><mi>c</mi><mi> </mi><mi>maximizes</mi><mi> </mi><mi>p</mi><mfenced><mi>c</mi></mfenced><mo>⁢</mo><mi>p</mi><mfenced><msub><mi>w</mi><mrow><mi>i</mi><mo>⁢</mo><mi>j</mi></mrow></msub><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mtd></mtr><mtr><mtd><mn>0</mn></mtd><mtd><mi>otherwise</mi></mtd></mtr></mtable></mrow></mtd></mtr></mtable></math><img file="EP1669980B1_D0018.tif" /></maths>

0075The second option may be undesirably slow for all but very small N-best lists.

D. Merging of Clusters

0076At this point, some of the clusters may represent similar semantic intents. Therefore, it may be desirable to merge some of the clusters together based on a distance measure between two clusters (or between the representative language models of the two clusters), and to refine the merged clusters. This is indicated by blocks 314 and 316 in <figref idref="f0004">FIG. 4</figref>.

0077Techniques for merging and splitting clustered items have been studied in the field of text clustering. Many of these techniques are based on certain distance measures between two clusters. In one embodiment, the present invention uses a relatively low complexity distance measure based on the K-L divergence between the uni-gram distributions corresponding to two clusters. K-L divergence is explained in greater detail in <nplcit id="ncit0002" npl-type="b"><text>T.M. Cover and J.A. Thomas, Elements of Information Theory, Wiley, (1991</text></nplcit>).

0078Assuming γ_{<i>c</i>,<i>u</i>} is the uni-gram probability of vocabulary item <i>u</i> in cluster <i>c</i> (γ_{<i>c</i>,<i>u</i>} is proportional to Φ<i>_c,u</i>), the distance is defined as an average of the asymmetrical K-L divergences, <maths id="math0019" num="Eq. 19"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>19</mn></mtd></mtr><mtr><mtd><mi>D</mi><mfenced><msub><mi>c</mi><mn>1</mn></msub><msub><mi>c</mi><mn>2</mn></msub></mfenced><mo>≜</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>u</mi></munder><mo>⁢</mo><mfenced><msub><mi>γ</mi><mrow><mi>c</mi><mo>⁢</mo><mn>1</mn><mo>,</mo><mi>u</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>γ</mi><mrow><mi>c</mi><mo>⁢</mo><mn>2</mn><mo>,</mo><mi>u</mi></mrow></msub><msub><mi>γ</mi><mrow><mi>c</mi><mo>⁢</mo><mn>1</mn><mo>,</mo><mi>u</mi></mrow></msub></mfrac><mo>+</mo><msub><mi>γ</mi><mrow><mi>c</mi><mo>⁢</mo><mn>2</mn><mo>,</mo><mi>u</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>γ</mi><mrow><mi>c</mi><mo>⁢</mo><mn>1</mn><mo>,</mo><mi>u</mi></mrow></msub><msub><mi>γ</mi><mrow><mi>c</mi><mo>⁢</mo><mn>2</mn><mo>,</mo><mi>u</mi></mrow></msub></mfrac></mfenced></mstyle><mo>,</mo></mtd></mtr></mtable></math><img file="EP1669980B1_D0019.tif" /></maths> where <i>u</i> is summed over all vocabulary items appearing in clusters <i>c₁</i> and <i>c₂</i>, and any zero probabilities γ_{<i>c</i>1,<i>u</i>} or <i>γ</i>_{<i>c</i>2,<i>u</i>} are smoothed by a floor value. Two clusters c₁ and c₂ are merged if their <i>D(c</i>₁<i>,c</i>₂<i>)</i> is smaller than a threshold. Upon merging, <i>p(w,c_1,2</i>|<i>x)=p(w,c₁</i>|<i>x)+p(w,c₂</i>|<i>x)</i> and the new model is re-estimated using these new posterior probabilities. A desired number of EM or Viterbi estimations are applied after all such pairs are merged.

0079In another embodiment, re-estimation can be applied after each pair is merged (the pair with the smallest divergence is to be merged first). But this can greatly increase computation and may thus be undesirable.

0080As an alternative to K-L divergence, the distance measure between two clusters and hence the measure to determine whether merging of two clusters should take place can be based on the EM auxiliary function. Specifically, the loss in the EM auxiliary function due to merging two clusters can be used as the distance measure between those clusters.

0081The EM auxiliary function is: <maths id="math0020" num="Eq. 20"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>20</mn></mtd></mtr><mtr><mtd><mstyle displaystyle="true"><munder><mo>∑</mo><mi>i</mi></munder><mstyle displaystyle="true"><munder><mo>∑</mo><mrow><mi>c</mi><mo>,</mo><mi>w</mi></mrow></munder><mo>⁢</mo><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mi>log</mi><mo>⁢</mo><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>c</mi></mfenced><mo>⁢</mo><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>w</mi><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mstyle></mstyle><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>M</mi></munderover><mstyle displaystyle="true"><munder><mo>∑</mo><mrow><mi>c</mi><mo>,</mo><mi>w</mi></mrow></munder><mfenced><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced></mfenced><mo>⁢</mo><mfenced open="[" close="]"><mi>log</mi><mo>⁢</mo><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>c</mi></mfenced><mo>+</mo><mi>log</mi><mo>⁢</mo><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>w</mi><mrow><mo>|</mo><mi>c</mi></mrow></mfenced></mfenced></mstyle></mstyle><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>M</mi></munderover><mstyle displaystyle="true"><munder><mo>∑</mo><mrow><mi>c</mi><mo>,</mo><mi>w</mi></mrow></munder><mo>⁢</mo><mi>p</mi><mfenced><mi>c</mi><mo>,</mo><mi>w</mi><mrow><mo>|</mo><msub><mi>x</mi><mi>i</mi></msub></mrow></mfenced><mo>⁢</mo><mfenced open="[" close="]"><mi>log</mi><mo>⁢</mo><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>c</mi></mfenced><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><mo>⁢</mo><msub><mo>#</mo><mi>v</mi></msub><mfenced><mi>w</mi></mfenced><mo>⁢</mo><mi>log</mi><mo>⁢</mo><msub><msup><mi>p</mi><mo>′</mo></msup><mi>c</mi></msub><mfenced><mi>v</mi></mfenced></mstyle></mfenced></mstyle></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0020.tif" /></maths>

0082Using count definitions from the E-step in the above-described EM algorithm: <maths id="math0021" num="Eq. 21"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>21</mn></mtd></mtr><mtr><mtd><mo>=</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>c</mi></munder><mo>⁢</mo><msub><mi>ψ</mi><mi>c</mi></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><msup><mi>p</mi><mo>′</mo></msup><mfenced><mi>c</mi></mfenced></mstyle><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mrow><mi>c</mi><mo>,</mo><mi>v</mi></mrow></munder><mo>⁢</mo><msub><mi>ψ</mi><mrow><mi>c</mi><mo>⁢</mo><mi>v</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><msub><msup><mi>p</mi><mo>′</mo></msup><mi>c</mi></msub><mfenced><mi>v</mi></mfenced></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0021.tif" /></maths> using <i>p'(c)</i> and <i>p'_c(v)</i> from the M-step described above: <maths id="math0022" num="Eq. 22"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>22</mn></mtd></mtr><mtr><mtd><mo>=</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>c</mi></munder><mo>⁢</mo><msub><mi>ψ</mi><mi>c</mi></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><mi>c</mi></msub><mi>M</mi></mfrac></mstyle><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mrow><mi>c</mi><mo>,</mo><mi>v</mi></mrow></munder><mo>⁢</mo><msub><mi>ψ</mi><mrow><mi>c</mi><mo>⁢</mo><mi>v</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><mrow><mi>c</mi><mo>⁢</mo><mi>v</mi></mrow></msub><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>v</mi><mo>′</mo></msup></munder><msub><mi>ψ</mi><mrow><mi>c</mi><mo>⁢</mo><msup><mi>v</mi><mo>′</mo></msup></mrow></msub></mstyle></mfrac></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0022.tif" /></maths>

0083If we're considering clusters <i>c₁</i> and <i>c₂</i>, then the unmerged auxiliary function is computed as: <maths id="math0023" num="Eq. 23"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>23</mn></mtd></mtr><mtr><mtd><msub><mi>Aux</mi><mi>unmerged</mi></msub><mo>=</mo><mfenced><mi>the value independent of</mi><mi> </mi><msub><mi>c</mi><mn>1</mn></msub><mo>,</mo><msub><mi>c</mi><mn>2</mn></msub></mfenced><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mi>M</mi></mfrac><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><mo>⁢</mo><msub><mi>ψ</mi><mrow><mi>c</mi><mo>,</mo><mi>v</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><mrow><mi>c</mi><mo>,</mo><mi>v</mi></mrow></msub><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>v</mi><mo>′</mo></msup></munder><msub><mi>ψ</mi><mrow><mi>c</mi><mo>,</mo><msup><mi>v</mi><mo>′</mo></msup></mrow></msub></mstyle></mfrac></mstyle><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub><mi>M</mi></mfrac><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><mo>⁢</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>v</mi><mo>′</mo></msup></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle></mfrac></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0023.tif" /></maths> If <i>c₁</i> and <i>c₂</i> are merged, the merged auxiliary function is computed as follows: <maths id="math0024" num="Eq. 24"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>24</mn></mtd></mtr><mtr><mtd><msub><mi>Aux</mi><mi>merge</mi></msub><mo>=</mo><mfenced><mi>the value independent of</mi><mi> </mi><msub><mi>c</mi><mn>1</mn></msub><mo>,</mo><msub><mi>c</mi><mn>2</mn></msub></mfenced><mo>+</mo><mfenced><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub></mfenced><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><mrow><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub></mrow><mi>M</mi></mfrac><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><mo>⁢</mo><mfenced><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>+</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mfenced><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><mrow><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub></mrow><mstyle displaystyle="true"><munder><mo>∑</mo><msup><mi>v</mi><mo>′</mo></msup></munder><mfenced><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><msup><mi>v</mi><mo>′</mo></msup></mrow></msub><mo>+</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><msup><mi>v</mi><mo>′</mo></msup></mrow></msub></mfenced></mstyle></mfrac></mstyle></mtd></mtr></mtable></math><img file="EP1669980B1_D0024.tif" /></maths>

0084The distance between <i>c₁</i> and <i>c₂</i> can be defined as the difference: <maths id="math0025" num="Eq. 25"><math display="block"><mtable columnalign="left"><mtr><mtd><mi>Eq</mi><mo>.</mo><mi> </mi><mn>25</mn></mtd></mtr><mtr><mtd><mi mathvariant="normal">D</mi><mfenced><msub><mi mathvariant="normal">c</mi><mn>1</mn></msub><msub><mi mathvariant="normal">c</mi><mn>2</mn></msub></mfenced><mo>=</mo><msub><mi>Aux</mi><mi>unmerged</mi></msub><mo>−</mo><msub><mi>Aux</mi><mi>merged</mi></msub><mo>=</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mrow><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub></mrow></mfrac><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub><mrow><msub><mi>ψ</mi><msub><mi>c</mi><mn>1</mn></msub></msub><mo>+</mo><msub><mi>ψ</mi><msub><mi>c</mi><mn>2</mn></msub></msub></mrow></mfrac><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><mo>⁢</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>⁢</mo><mfrac><mrow><mi>log</mi><mo>⁢</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mrow><mrow><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>+</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mrow></mfrac></mstyle><mo>+</mo><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><mo>⁢</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mrow><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>+</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mrow></mfrac></mstyle><mo>−</mo><mfenced><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle></mfenced><mo>⁢</mo><mi>log</mi><mo>⁢</mo><mfrac><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>+</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle></mfrac><mo>−</mo><mfrac><mrow><mfenced><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle></mfenced><mo>⁢</mo><mi>log</mi><mfenced><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle></mfenced></mrow><mfenced><mstyle displaystyle="true"><munder><mo>∑</mo><mi>v</mi></munder><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>1</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub><mo>+</mo><msub><mi>ψ</mi><mrow><msub><mi>c</mi><mn>2</mn></msub><mo>⁢</mo><mi>v</mi></mrow></msub></mstyle></mfenced></mfrac></mtd></mtr></mtable></math><img file="EP1669980B1_D0025.tif" /></maths>

0085The loss of perplexity can also be used in determining the distance between two clusters. Perplexity is described in greater detail in the following papers: <nplcit id="ncit0003" npl-type="s"><text>Young, Odell and Woodland, Tree-Based State Tying for High Accuracy Acoustic Modeling, ARPA, pages 307-312 (March 1994</text></nplcit>); and <nplcit id="ncit0004" npl-type="s"><text>Hwang and Huang, Shared-Distribution Hidden Markov Models for Speech Recognition, IEEE TSAP, volume 1, number 4, pages 414-420 (1993)</text></nplcit>.

0086In another embodiment, the clusters are merged until the auxiliary function changes by some predetermined amount rather than merging all clusters with a distance less than a threshold. The amount of change in the auxiliary function used to determine whether clusters are merged can be a predetermined percentage or other value empirically determined.

0087Further, the merging process can be repeated a plurality of times, interspersed with re-estimation. This is referred to as iterative merging.

0088Recall that once the similar clusters are merged, a representative utterance or label for each cluster is chosen. In one embodiment, this is based on the likelihood p(w|c) calculated for each utterance in each cluster. It should be noted that when the auxiliary function is used for merging and this likelihood is used for choosing a cluster representative, then merging, re-estimation and representative selection are all consistent (performed using the same criteria) and the implementation may thus be simpler.

0089It can thus be seen that the present invention provides significant advantages over prior systems. For instance, the present invention automatically clusters acoustics based on semantic intent. The present invention can also identify a representative acoustic record (or speech recognition record) representative of each cluster. Therefore, a developer need not listen to a large amount of data to identify unanticipated semantic intents in order to adapt application grammars to accommodate those semantic intents.

0090The present invention can also be used to suggest grammar rules or models to modify the application grammars either manually or automatically. The present invention also provides significant advantages in how it extracts data, clusters that data based on speech recognition results corresponding to that data, and trains representative models, representative of each cluster.

0091Although 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 scope of the invention.