Speech recognition based on a multilingual acoustic model

Field of Invention

0001The present invention relates to the art of automatic speech recognition and, in particular, the generation of an acoustic model for speech recognition of spoken utterances in different languages.

Background of the Invention

0002The human voice can probably be considered as the most natural and comfortable man-computer interface. Voice input provides the advantages of hands-free operation, thereby, e.g., providing access for physically challenged users or users that are using there hands for different operation, e.g., driving a car. Thus, computer users for a long time desired software applications that can be operated by verbal utterances.

0003During speech recognition verbal utterances, either isolated words or continuous speech, are captured by a microphone or a telephone, for example, and converted to analogue electronic signals that subsequently are digitized. The digital signals are usually subject to a subsequent spectral analysis. Recent representations of the speech waveforms sampled typically at a rate between 6.6 kHz and 20 kHz are derived from the short term power spectra and represent a sequence of characterizing vectors containing values of what is generally referred to as features/feature parameters. The values of the feature parameters are used in succeeding stages in the estimation of the probability that the portion of the analyzed waveform corresponds to, for example, a particular entry, i.e. a word, in a vocabulary list.

0004Present-day speech recognition systems usually make use of acoustic and language models. The acoustic models comprise codebooks consisting of Gaussians representing typical sounds of human speech and Hidden Markov Models (HMMs). The HMMs represent allophones/phonemes a concatenation of which constitute a linguistic word. The HMMs are characterized by a sequence of states each of which has a well-defined transition probability. In order to recognize a spoken word, the systems have to compute the most likely sequence of states through the HMM. This calculation is usually performed by means of the Viterbi algorithm, which iteratively determines the most likely path through the associated trellis. The language model, on the other hand, describes the probabilities of sequences of words and/or a particular grammar.

0005The reliability of the correct speech recognition of a verbal utterance of an operator is a main task in the art of speech recognition/operation and despite recent progress still raises demanding problems, in particular, in the context of embedded systems that suffer from severe memory and processor limitations. These problems are eminently considerable when speech inputs of different languages are to be expected. A driver of car, say a German mother-tongue driver, might need to input an expression, e.g., representing a town, in a foreign language, say in English. To give another example, different native users of an MP3/MP4 player or a similar audio device will assign tags in different languages. Furthermore, titles of songs stored in the player may be of different languages (e.g., English, French, German).

0006<patcit id="pcit0001" dnum="US20070294082A"><text>US 2007/0294082</text></patcit> and a paper by <nplcit id="ncit0001" npl-type="s"><text>M. Raab, R. Gruhn and E. Nöth, entitled "Multilingual Weighted Codebooks for Non-Native Speech Recognition", Proceedings of the 11th International Conference on Text, Speech and Dialogue, TSD 2008, vol. 5426/2008, pages 485-492</text></nplcit>, disclose methods for the generation of a multilingual recognizer by increasing the number of possible states/models of vocal units in a recognizer by adding states/models of vocal units appropriate for a language other than the language for which the recognizer was initially trained

0007<nplcit id="ncit0002" npl-type="s"><text>G. Stemmer et al., in a paper entitled "Acoustic Modeling of Foreign Words in a German Speech Recognition System", Proceedings of Eurospeech 2001, vol. 4, pages 2745-2748</text></nplcit>, describe a method for the acoustic modeling of foreign words in a German recognizer wherein foreign phonemes are mapped to German ones or a merging of German and foreign phonemes is performed.

0008<nplcit id="ncit0003" npl-type="s"><text>J. Köhler, in a paper entitled "Multi-Lingual Phoneme Recognition exploiting acoustic-phonetic similarities of sounds", Proceedings of the international conference on spoken language processing, 3 October 1996, pages 2195-2198</text></nplcit>, describe multi-lingual English-German speech recognition without substitution of German phonemes with the corresponding English phonemes within multilingual phoneme recognition.

0009Present day speech recognition and control means usually comprise codebooks that are commonly generated by the (generalized) Linde-Buzo-Gray (LBG) algorithm or related algorithms. However, such kind of codebook generation aims to find a limited number of (Gaussian) prototype code vectors in the feature space covering the entire training data which usually comprises data of one single language. Moreover, in conventional multilingual applications all Gaussians of multiples codebooks generated for different languages have to be searched during a recognition process. In particular, in embedded systems characterized by rather limited computational resources this can result in an inconvenient or even unacceptable processing time. In addition, when a new language has to be recognized that is not already considered by a particular speech recognition means exhaustive training on new speech data has to be performed which is not achievable by embedded system with limited memory and processor power.

0010Thus, there is a need for reliable and fast speech recognition of speech inputs of different languages that, in particular, is acceptable in terms of the demands for computer resources.

Description of the invention

0011In view of the above, the present invention provides a method for generating a multilingual speech recognizer comprising a multilingual acoustic model, comprising the steps of providing a first speech recognizer comprising a first codebook consisting of first Gaussians and first Hidden Markov Models, HMMs, comprising first states; providing at least one second speech recognizer comprising a second codebook consisting of second Gaussians and second Hidden Markov Models, HMMs, comprising second states; replacing each of the second Gaussians of the at least one second speech recognizer by the respective closest one of the first Gaussians and/or each of the second states of the second HMMs of the at least one second speech recognizer with the respective closest state of the first HMMs of the first speech recognizer to obtain at least one modified second speech recognizer; and combining the first speech recognizer and the at least one modified second speech recognizer to obtain the multilingual speech recognizer.

0012The first and the second speech recognizers are each trained for a different language based on speech data as known in the art. The speech data is usually provided by one or more respective native speakers. For example, the first speech recognizer may be configured to recognize speech inputs in English and the second one may be configured to recognize speech inputs in French or German. According to the present invention the first speech recognizer is not modified in order to generate a multilingual speech recognizer. The language recognized by the first speech recognizer may be considered the main language, e.g., the language of a native user of a device incorporating the present invention. More than one second speech recognizers configured for different languages other than the main language can be combined after the above-mentioned modification(s) with the unmodified first speech recognizer in order to obtain the multilingual speech recognizer based on a multilingual acoustic model comprising Gaussians of the first codebook only (according to this embodiment) for both the recognition of a language for which the first codebook was provided and a different language for which the second codebook was provided.

0013Each of the speech recognizers is conventionally configured to recognize a speech input based on <ol id="ol0001" ol-style=""><li>a) an acoustic model comprising a codebook consisting of Gaussians and a trained Hidden Markov Model (HMM) comprising states; and</li><li>b) a language (grammar) model describing the probability/allowance of consecutive words and/or sentences in a particular language.</li></ol>

0014The Gaussians represent well-known probability distributions describing typical sounds of human speech in a particular language. The Gaussians may be considered as vectors, i.e. a Gaussian density distribution of feature vectors related to features as the pitch, spectral envelope, etc. and generated for each language by some technique as known in the art. The HMMs produce likelihoods of sequences of single speech fragments represented by the Gaussians. In particular, the HMMs may represent phonemes or allophones. The actual pronunciation of a word of a language can be represented as an HMM sequence.

0015When a single word is recognized by one of the speech recognizers (be it the first recognizer, the at least one second recognizer or the multilingual speech recognizer generated in accordance with the present invention), the recognition result is the sequence of HMMs that produces the highest overall likelihood of all HMM sequences allowed by the language model, for example. In some more detail, the HMMs might consider all Gaussians according to weights being part of the acoustic model and saved in so-called "B-matrices". An HMM is split into several states each of which having a separate B-matrix (see also detailed description below).

0016The computational costs of speech recognition in terms of processor load, processing speed and memory demand depends on the number of different active HMM states and the number of Gaussians. The generation of the multilingual speech recognizer from multiple conventionally trained monolingual according to the herein disclosed method can readily be performed on-the-fly even on embedded systems with restricted computational resources. New languages can easily be added to an existing multilingual speech recognizer.

0017According to the above-described embodiment of the inventive method for generating a multilingual speech recognizer for speech recognition of speech inputs in different languages the multilingual speech recognizer is generated by comprising/maintaining not all Gaussians of two or more speech recognizers provided for different languages but only Gaussians of the first speech recognizer configured for recognition of a first/main/mother language. In fact, according to this embodiment not one single Gaussian from a speech recognizer configured for recognition of a speech input in a language different from the main language is comprised in the multilingual speech recognizer. The number of possible active HMM states in the achieved multilingual speech recognizer remains the same as in the first speech recognizer provided for the main language.

0018The resulting multilingual speech recognizer can be used for fast and relatively reliable multilingual speech recognition of speech inputs of different languages in the embedded systems in a reasonable processing time. Since the first codebook was maintained without modification the multilingual speech recognizer works optimally for the corresponding main language usually used by a native speaker and worse for other languages. However, experimental studies have proven that in most cases the overall performance of the multilingual speech recognizer is acceptable for speech inputs in languages other than the main language.

0019Alternatively or supplementary to replacing the second Gaussians of the second codebook of the at least one second speech recognizer by the respective closest first Gaussians of the unmodified first codebook of the first speech recognizer the states of the second HMM of the at least one second speech recognizer are replaced with the respective closest states of the first HMM of the first speech recognizer to obtain at least one modified second speech recognizer an, eventually, the desired multilingual speech recognizer.

0020Whereas, in principle, any kind of distance measure known in the art can be used to determine the respective closest Gaussians or states, according to an example the closest Gaussians are determined based on the Mahalanobis distance. This distance measure provides an efficient means for determining closest Gaussians and results in a successful multilingual speech recognizer. Closest states of the HMMs, on the other hand, may be determined based on the Euclidian distances of the states of the second HMM to the states of the first HMM, i.e. states from different languages (see detailed description below).

0021Alternatively or supplementary to replacing Gaussians and/or states of HMMs the second HMM of the at least one second speech recognizer may be replaced by the closest HMM of the first speech recognizer in order to obtain the multilingual speech recognizer. The closest HMM can for example be determined by the minimum sum of distances between the states of two HMMs. Another possibility would be to compare the expected values of HMMs.

0022Thus, it is provided a method for generating a speech recognizer comprising a multilingual acoustic model; comprising the steps of providing a first speech recognizer comprising a first codebook consisting of first Gaussians and a first Hidden Markov Model, HMM, comprising first states; providing at least one second speech recognizer comprising a second codebook consisting of second Gaussians and a second Hidden Markov Model, HMM, comprising second states; determining mean vectors of states for the first states of the first HMM of the first speech recognizer; determining HMMs of the first speech recognizer based on the determined mean vectors of states; replacing the second HMM of the at least one second speech recognizer by the closest HMM of the first speech recognizer (by the particular HMM of first speech recognizer determined based on the determined mean vectors of states that is closest to the HMM of the second speech recognizer) to obtain at least one modified second speech recognizer; and combining the first speech recognizer and the at least one modified second speech recognizer to obtain the multilingual speech recognizer.

0023In this example, the at least one modified second speech recognizer may be obtained by also replacing each of the second Gaussians of the at least one second speech recognizer by the respective closest one of the first Gaussians and/or each of the second states of the second HMM of the at least one second speech recognizer with the respective closest state of the first HMM of the first speech recognizer to obtain the at least one modified second speech recognizer.

0024According to examples of this method two or three modified second speech recognizers are generated by means of the above-described procedures of replacing Gaussians, states of HMMs and HMMs by the first Gaussians of the first codebook, the states of the first HMM and the HMMs of the first speech recognizer generated from mean vectors of states for the first states of the first HMM of the first speech recognizer, respectively, and the multilingual speech recognizer is obtained by weighting the two or three modified second speech recognizers and subsequently combining the weighted modified second speech recognizers with the first speech recognizer.

0025For example, a first modified second speech recognizer may be generated by replacement of the Gaussians and a second modified second speech recognizer by replacement of HMM states as described above and the first (unmodified) speech recognizer and the first modified second speech recognizer weighted by a first weight (e.g., chosen from 0.4 to 0.6) and the second modified second speech recognizer weighted by a second weight (e.g., chosen from 0.4 to 0.6) are combined with each other to obtain the multilingual speech recognizer.

0026According to another example, a first modified second speech recognizer may be generated by replacement of the Gaussians, a second modified second speech recognizer by replacement of HMM states and a third modified second speech recognizer by replacement of the second HMM as described above and the first (unmodified) speech recognizer and the first modified second speech recognizer weighted by a first weight, the second modified second speech recognizer weighted by a second weight and the third modified second speech recognizer weighted by a second weight are combined with each other to obtain the multilingual speech recognizer. Adjustment of the weights may facilitate fine-tuning of the achieved multilingual speech recognizer and assist in improving the reliability of recognition results of speech inputs in different languages in different actual applications.

0027As already mentioned above, speech inputs in the main language for which the first speech recognizer is trained are recognized by means of the achieved multilingual speech recognizer with the same reliability as with the first speech recognizer. Recognition results for speech inputs in other languages tend to be worse. In view of this, according to an embodiment the first speech recognizer is modified by modifying the first codebook before combining it with the at least one modified second speech recognizer to obtain the multilingual speech recognizer, wherein the step of modifying the first codebook comprises adding at least one of the second Gaussians of the second codebook of the at least one second speech recognizer to the first codebook. Thereby, recognition results for speech inputs in a language other than the main language for which the first speech recognizer is trained are improved.

0028Advantageously, such Gaussians of the second codebook of the at least one second speech recognizer are added to the codebook of the generated multilingual speech recognizer that are very different from the first Gaussians of the first codebook. In particular, a sub-set of the second Gaussians of the second codebook is added to the first codebook based on distances between the second and the first Gaussians. In this case, the distances between the second and the first Gaussians are determined and at least one of the second Gaussians is added to the first codebook that exhibits a predetermined distance from one of the first Gaussians that is closest to this particular at least one of the second Gaussians.

0029The distance can be determined by means of the Mahalanobis distance or the Kullback-Leibler divergence or by minimizing the gain in variance when a particular additional code vector is merged with different particular code vectors of the main language codebook, i.e. when the respective (merging) code vectors are replaced by a code vector that would have been estimated from the training samples of both the main language codebook and the additional codebook that resulted in the code vectors that are merged. It is noted that based on experiments performed by the inventors the Mahalanobis distance has been proven to be a very suitable measure in this context.

0030By means of the multilingual speech recognizer according to one of the above-described examples speech recognition of speech inputs in different languages can be performed even in embedded systems with restricted computational resources. Thus, it is provided a method for speech recognition comprising speech recognition based on a multilingual speech recognizer provided by a method of one of the above-described examples. Speech recognition can be realized by a speech recognition means or speech dialog system or speech control system comprising a multilingual speech recognizer generated by the method according to one of the above-described examples.

0031Herein, it is further provided an audio device, in particular, an MP3 or MP4 player, a cell phone or a Personal Digital Assistant, or a video device comprising a speech recognition or speech dialog system or speech control system means comprising a multilingual speech recognizer generated according to the method according to one of the above-described examples.

0032Furthermore, it is provided a computer program product, comprising one or more computer readable media having computer-executable instructions for performing the steps of the method according to one of the above-described examples.

0033Additional features and advantages of the present invention will be described with reference to the drawing. In the description, reference is made to the accompanying figure that is meant to illustrate an example of the invention. It is understood that such an example does not represent the full scope of the invention.

0034<figref idref="f0001">Figure 1</figref> illustrates an example of the inventive method of generating a multilingual speech recognizer based on a multilingual acoustic model.

0035In the following, an example for the creation of a multilingual speech recognizer based on a multilingual acoustic model (multilingual HMMs) from a number of monolingual speech recognizer/acoustic models according to the present invention is described with reference to <figref idref="f0001">Figure 1</figref>. Multilingual HMMs are created by mapping the HMMs of speech recognizers provided for different languages to the Gaussians of one predetermined speech recognizer provided for another, main language. Consider an example of n speech recognizers 1 provided for n different languages and indicated by reference numerals 1, 2, 3 and 4 in <figref idref="f0001">Figure 1</figref>. It goes without saying that different from the shown example more than four speech recognizers in total can be employed.

0036Each speech recognizer comprises a language model and an acoustic model as known in the art. The respective acoustic models comprise Gaussians corresponding to speech fragments of the respective languages and organized in codebooks as well as HMMs representing phonemes. Each HMM model of each speech recognizer considers the likelihoods the Gaussians produce and adds additional likelihoods for accounting for changing from one HMM to another. Recognition of a particular spoken word provides a recognition result representing a sequence of HMM models giving the highest overall likelihood of all HMM sequences allowed according to the language model. Moreover, each HMM is split into a predetermined number of (HMM/language) states each of which is linked to a different B matrix including weights associated with the respective Gaussians.

0037According to the present example, the speech recognizer indicated by the reference number 4 corresponds to a language that is considered the main (native) language and recognition of utterances in that main language shall not be affected when performed based on the new multilingual speech recognizer that is to be created. In the following, the set of Gaussians of speech recognizer 4 is also denoted as recognition codebook. All Gaussians of all the other n-1 speech recognizers 1, 2 and 3 are mapped 5 to the Gaussians of the speech recognizer 4 as described in the following.

0038Each Gaussian is characterized by its mean µ and covariance matrix Σ. In this example, mapping 5 is based on the well-known Mahalanobis distance measure: <maths id="math0001" num=""><math display="block"><msub><mi>map</mi><mi mathvariant="normal">G</mi></msub><mfenced><msub><msup><mi mathvariant="normal">G</mi><mi mathvariant="normal">i</mi></msup><mi>MC</mi></msub></mfenced><mo>=</mo><msub><msup><mi mathvariant="normal">G</mi><mi mathvariant="normal">i</mi></msup><mi>RC</mi></msub><mo>,</mo><mn mathvariant="normal">0</mn><mo>≤</mo><mi mathvariant="normal">i</mi><mo>≤</mo><mi mathvariant="normal">M</mi><mo>,</mo><mn mathvariant="normal">0</mn><mo>≤</mo><mi mathvariant="normal">j</mi><mo>≤</mo><mi mathvariant="normal">N</mi></math><img file="EP2192575B1_D0001.tif" /></maths> with <maths id="math0002" num=""><math display="block"><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">G</mi></msub><mfenced separators=""><msub><msup><mi mathvariant="normal">μ</mi><mi mathvariant="normal">i</mi></msup><mi>MC</mi></msub><mo>,</mo><msub><msup><mi mathvariant="normal">μ</mi><mi mathvariant="normal">j</mi></msup><mi>RC</mi></msub><mo>,</mo><mstyle displaystyle="false"><munderover><mo>∑</mo><mi>MC</mi><mi mathvariant="normal">i</mi></munderover></mstyle><mspace width="1em" /></mfenced><mo>≤</mo><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">G</mi></msub><mfenced separators=""><msub><msup><mi mathvariant="normal">μ</mi><mi mathvariant="normal">i</mi></msup><mi>MC</mi></msub><mo>,</mo><msub><msup><mi mathvariant="normal">μ</mi><mi mathvariant="normal">k</mi></msup><mi>RC</mi></msub><mo>,</mo><mstyle displaystyle="false"><munderover><mo>∑</mo><mi>MC</mi><mi mathvariant="normal">i</mi></munderover></mstyle><mspace width="1em" /></mfenced><mspace width="1em" /><mo>∀</mo><mi mathvariant="normal">k</mi><mo>,</mo><mspace width="1em" /><mi>with</mi><mspace width="1em" /><mn mathvariant="normal">1</mn><mo>≤</mo><mi mathvariant="normal">k</mi><mo>≤</mo><mi mathvariant="normal">N</mi><mo>,</mo><mi mathvariant="normal">k</mi><mo>≠</mo><mi mathvariant="normal">j</mi><mn mathvariant="normal">.</mn></math><img file="EP2192575B1_D0002.tif" /></maths>

0039In the above-used notation the indices i, j, k indicate the respective individual Gaussians and RC and MC denote the recognition codebook and the respective monolingual codebooks of the speech recognizers 1, 2, and 3. All states from HMMs corresponding to the language that corresponds to speech recognizer 4 map to Gaussians of the recognition codebook (of the speech recognizer 4) only. Thus, when all states s of all HMMs of all the other recognizers 1, 2 and 3 are mapped to the RS HMM/language states represented by speech recognizer 4 only Gaussians of the recognition codebook of recognizer 4 are used for any recognition process based on the multilingual speech recognizer that is to be generated.

0040The mapping 6 of HMM/language states of the HMMs, according to the present example, is performed based on the minimum Euclidean distance measure (D_Eu) between expected values of the probability distributions of the states. Here, it is assumed that the probability distribution p_s of every state s of an HMM is a Gaussian mixture distribution. It is, furthermore, assumed that all MCs have N Gaussians and each state has N weights w.

0041Then, the probability distribution p_s of every state s is given by <maths id="math0003" num=""><math display="block"><msub><mi mathvariant="normal">p</mi><mi mathvariant="normal">s</mi></msub><mfenced><mi mathvariant="normal">x</mi></mfenced><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi mathvariant="normal">i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mi mathvariant="normal">N</mi></munderover></mstyle><msub><mi mathvariant="normal">w</mi><mi mathvariant="normal">i</mi></msub><mo>*</mo><mi mathvariant="normal">G</mi><mfenced separators=""><mi mathvariant="normal">x</mi><mo>⁢</mo><msub><mi mathvariant="normal">μ</mi><mi mathvariant="normal">i</mi></msub><mstyle displaystyle="false"><munder><mo>∑</mo><mi mathvariant="normal">i</mi></munder></mstyle></mfenced><mn mathvariant="normal">.</mn></math><img file="EP2192575B1_D0003.tif" /></maths>

0042The expectation value for each state s can readily be obtained by <maths id="math0004" num=""><math display="block"><mi mathvariant="normal">E</mi><mfenced separators=""><msub><mi mathvariant="normal">p</mi><mi mathvariant="normal">s</mi></msub><mfenced><mi mathvariant="normal">x</mi></mfenced></mfenced><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi mathvariant="normal">i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mi mathvariant="normal">N</mi></munderover></mstyle><mi mathvariant="normal">E</mi><mfenced separators=""><msup><msub><mi mathvariant="normal">w</mi><mi mathvariant="normal">i</mi></msub><mi mathvariant="normal">s</mi></msup><mo>*</mo><mi mathvariant="normal">G</mi><mfenced separators=""><mi mathvariant="normal">x</mi><mo>⁢</mo><msub><mi mathvariant="normal">μ</mi><mi mathvariant="normal">i</mi></msub><mstyle displaystyle="false"><munder><mo>∑</mo><mi mathvariant="normal">i</mi></munder><mspace width="1em" /></mstyle></mfenced></mfenced><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi mathvariant="normal">i</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow><mi mathvariant="normal">N</mi></munderover></mstyle><msup><msub><mi mathvariant="normal">w</mi><mi mathvariant="normal">i</mi></msub><mi mathvariant="normal">s</mi></msup><mo>*</mo><msub><mi mathvariant="normal">μ</mi><mi mathvariant="normal">i</mi></msub><mn mathvariant="normal">.</mn></math><img file="EP2192575B1_D0004.tif" /></maths>

0043Thus, the distance D_S between two particular states s¹ and s² can be defined by <maths id="math0005" num=""><math display="block"><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">s</mi></msub><mfenced separators=""><msup><mi mathvariant="normal">s</mi><mn mathvariant="normal">1</mn></msup><mo>⁢</mo><msup><mi mathvariant="normal">s</mi><mn mathvariant="normal">2</mn></msup></mfenced><mo>=</mo><msub><mi mathvariant="normal">D</mi><mi>Eu</mi></msub><mspace width="1em" /><mfenced separators=""><mi mathvariant="normal">E</mi><mfenced><msub><mi mathvariant="normal">p</mi><msup><mi mathvariant="normal">s</mi><mn mathvariant="normal">1</mn></msup></msub></mfenced><mo>,</mo><mi mathvariant="normal">E</mi><mfenced><msub><mi mathvariant="normal">p</mi><msup><mi mathvariant="normal">s</mi><mn mathvariant="normal">2</mn></msup></msub></mfenced></mfenced><mn mathvariant="normal">.</mn></math><img file="EP2192575B1_D0005.tif" /></maths>

0044In the present example, each speech recognizer has its own Linear Discriminant Analysis (LDA) transformation. Since the above-equation is correct only, if all states refer to Gaussians in the same feature space (language), the LDA is reversed before calculation of the distance D_S between states from different languages.

0045With the distance D_S as given above the state based mapping can be performed according to <maths id="math0006" num=""><math display="block"><msub><mi>map</mi><mi mathvariant="normal">s</mi></msub><mfenced><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup></mfenced><mo>=</mo><msub><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">j</mi></msup><mi>RS</mi></msub><mo>,</mo><mn mathvariant="normal">0</mn><mo>≤</mo><mi mathvariant="normal">i</mi><mo>≤</mo><mi mathvariant="normal">S</mi><mo>,</mo><mn mathvariant="normal">0</mn><mo>≤</mo><mi mathvariant="normal">j</mi><mo>≤</mo><mi>RS</mi></math><img file="EP2192575B1_D0006.tif" /></maths><maths id="math0007" num=""><math display="block"><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">s</mi></msub><mfenced separators=""><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup><mo>⁢</mo><msub><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">j</mi></msup><mi>RS</mi></msub></mfenced><mo>≤</mo><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">s</mi></msub><mfenced separators=""><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup><mo>⁢</mo><msub><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">k</mi></msup><mi>RS</mi></msub></mfenced><mspace width="1em" /><mo>∀</mo><mi mathvariant="normal">k</mi><mo>,</mo><mspace width="1em" /><mi>with</mi><mspace width="1em" /><mn mathvariant="normal">1</mn><mo>≤</mo><mi mathvariant="normal">k</mi><mo>≤</mo><mi>RS</mi><mo>,</mo><mi mathvariant="normal">k</mi><mo>≠</mo><mi mathvariant="normal">j</mi><mn mathvariant="normal">.</mn></math><img file="EP2192575B1_D0007.tif" /></maths>

0046Based on the distances between the states of the HMMs a distance between entire HMMs can be calculated. If, for example, each (context dependent) phoneme is represented by a three state HMM, the distance between two phonemes q₁ and q₂ is given by <maths id="math0008" num=""><math display="block"><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">H</mi></msub><mfenced separators=""><msub><mi mathvariant="normal">q</mi><mn mathvariant="normal">1</mn></msub><mo>⁢</mo><msub><mi mathvariant="normal">q</mi><mn mathvariant="normal">2</mn></msub></mfenced><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi mathvariant="normal">i</mi><mo>=</mo><mn mathvariant="normal">1</mn></mrow><mn mathvariant="normal">3</mn></munderover></mstyle><msub><mi mathvariant="normal">D</mi><mi mathvariant="normal">S</mi></msub><mfenced separators=""><msub><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup><msub><mi mathvariant="normal">q</mi><mn mathvariant="normal">1</mn></msub></msub><mo>⁢</mo><msub><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup><msub><mi mathvariant="normal">q</mi><mn mathvariant="normal">2</mn></msub></msub></mfenced><mn mathvariant="normal">.</mn></math><img file="EP2192575B1_D0008.tif" /></maths>

0047Similar to the mapping of Gaussians 5 and the state based mapping 6 described above, HMM mapping 7 from the HMMs of speech recognizers 1, 2 and 3 to speech recognizer 4 can be performed. According to the present invention, one of the three kinds of mapping or any combination of these mappings can be used in order to achieve the desired multilingual model.

0048Experimental studies have shown that, e.g., a combination of the mapping of Gaussians and HMM states at equal weights results in a reliable multilingual speech recognizer. A combined mapping of Gaussians and states of HMMs based on D_G and D_S can realized by <maths id="math0009" num=""><math display="block"><msub><mi>map</mi><mrow><mi mathvariant="normal">G</mi><mo>+</mo><mi mathvariant="normal">S</mi></mrow></msub><mfenced><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup></mfenced><mo>=</mo><msub><mi mathvariant="normal">Y</mi><mrow><mi mathvariant="normal">G</mi><mo>+</mo><mi mathvariant="normal">S</mi></mrow></msub><mspace width="1em" /><msub><mi>map</mi><mi mathvariant="normal">S</mi></msub><mfenced><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup></mfenced><mo>+</mo><mfenced separators=""><mn mathvariant="normal">1</mn><mo>-</mo><msub><mi mathvariant="normal">y</mi><mrow><mi mathvariant="normal">G</mi><mo>+</mo><mi mathvariant="normal">S</mi></mrow></msub></mfenced><mo>⁢</mo><mfenced><mtable><mtr><mtd><msub><msup><mi mathvariant="normal">w</mi><mn mathvariant="normal">1</mn></msup><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup></msub></mtd><mtd><msub><mi>map</mi><mi mathvariant="normal">G</mi></msub><mfenced><msub><msup><mi mathvariant="normal">G</mi><mn mathvariant="normal">1</mn></msup><mi>MC</mi></msub></mfenced></mtd></mtr><mtr><mtd><msub><msup><mi mathvariant="normal">w</mi><mn mathvariant="normal">1</mn></msup><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup></msub></mtd><mtd><msub><mi>map</mi><mi mathvariant="normal">G</mi></msub><mfenced><msub><msup><mi mathvariant="normal">G</mi><mn mathvariant="normal">2</mn></msup><mi>MC</mi></msub></mfenced></mtd></mtr><mtr><mtd><mspace width="1em" /></mtd><mtd><mo>⋮</mo></mtd></mtr><mtr><mtd><msub><msup><mi mathvariant="normal">w</mi><mi mathvariant="normal">N</mi></msup><msup><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi></msup></msub></mtd><mtd><msub><mi>map</mi><mi mathvariant="normal">G</mi></msub><mfenced><msub><msup><mi mathvariant="normal">G</mi><mi mathvariant="normal">N</mi></msup><mi>MC</mi></msub></mfenced></mtd></mtr></mtable></mfenced><mo>,</mo><mo>∀</mo><mspace width="1em" /><mi>i with</mi><mspace width="1em" /><mn mathvariant="normal">1</mn><mo>≤</mo><mi mathvariant="normal">i</mi><mo>≤</mo><mi mathvariant="normal">S</mi><mo>,</mo></math><img file="EP2192575B1_D0009.tif" /></maths> where γ_G+S is the weight of the combined mapping (γ_G+S = 0.5 for an equally weighted mapping). For a given application, the weight γ_G+S can be determined by experiments. In any case, no retraining of the resulting multilingual speech recognizer is necessary after the mapping process.