JP5411837B2 - Acoustic model creation device, acoustic model creation method, and program thereof - Google Patents

Description

  The present invention relates to an acoustic model creation device, an acoustic model creation method, and a program for creating an acoustic model mainly used in speech recognition processing.

  In speech recognition processing by a decoder (speech recognition engine or search processing unit of speech recognition engine), a phoneme model used between words mainly uses a phoneme environment, and is based on a WFST (weighted finite state transducer) base or the like. The 1-pass decoder uses a triphone that considers the phoneme environment before and after the central phoneme, whereas the 2-pass decoder uses a biphone that considers either the phoneme environment before or after the first pass or a monophone that does not consider the phoneme environment (hereinafter referred to as the phoneme environment). "Bi / monophone") may be used or triphone may be used, and generally the processing speed is faster when using bi / monophone (note that both use triphone in the second pass). When the phoneme model used between words is only triphone, learning is performed only with triphone learning data, but when there is a possibility of using bi / monophone, learning is also performed with bi / monophone learning data.

Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, Mikio Yamamoto, “IT Text Speech Recognition System”, Ohmsha, 2001, p.144

  In speech recognition processing, a decoder that uses only triphone between words (hereinafter referred to as “interphone triphone use decoder”) and a decoder that uses not only triphone but also bi / monophone between words (hereinafter “inter-word bi / monophone”). In the case of using the “decoder to be used”), it is necessary to separately prepare an acoustic model suitable for each decoder as shown in the apparatus configuration example of FIG. An acoustic model is a set of phoneme models including a necessary number of different phone models of triphone, biphone, and monophone.

  Specifically, for the inter-word triphone use decoder 10, an inter-word triphone use acoustic model storage unit 105 in which an acoustic model for learning only triphone (hereinafter referred to as “inter-word triphone use acoustic model”) is stored. The inter-word bi / monophone use decoder 20 stores an acoustic model for learning each of triphone, biphone, and monophone (hereinafter referred to as “inter-bi bi / monophone use acoustic model”). An inter-bi / monophone use acoustic model storage unit 110 is prepared (note that the language model 30 is also necessary, but the description is omitted because it is irrelevant to the present invention). This is due to the following reason.

  First, consider that an inter-word triphone use acoustic model for learning only triphone between words is shared by the inter-word triphone use decoder 10 and the inter-word bi / monophone use decoder 20. In this case, when a triphone (for example, k-a + k) phoneme model does not exist in the acoustic model, if the triphone is input as learning data, bi / monophone (with phoneme environment close to the triphone ( For example, a phoneme model of k-a + * or * -a + *) is learned with the learning data of the triphone (falling into such a situation is called “back-off”). For this reason, if the bi / monophone phoneme model is used as it is in the inter-word bi / monophone use decoder 20, the speech recognition accuracy deteriorates.

  Next, it is considered that the inter-word bi / monophone use acoustic model for learning triphone, biphone, and monophone between words is shared by the inter-word bi / monophone use decoder 20 and the inter-word triphone use decoder 10. Also in this case, when there is no phone model of a certain triphone (for example, k-a + k) in the acoustic model, when the triphone is input as learning data, bi / having a phoneme environment close to the triphone. A phone model of monophone (for example, k-a + *, * -a + *, * is an environment-independent symbol) is learned from the training data of the triphone by back-off. Since the bi / monophone phoneme model is also learned by bi / monophone learning data, the dispersion becomes wide and the discrimination performance deteriorates. Therefore, when the acoustic model is used in the inter-word triphone use decoder 10, the speech recognition accuracy is deteriorated as compared with the case where the inter-word triphone use acoustic model for learning only triphone between words is used.

  An object of the present invention is to create an acoustic model that can be shared by both decoders without impairing the voice recognition accuracy in each decoder when using an interword triphone use decoder and an interword bi / monophone use decoder. An acoustic model creation apparatus, an acoustic model creation method, and a program thereof are provided.

The acoustic model creation device of the present invention includes an inter-word bi / monophone use acoustic model storage unit, an inter-word tri / bi / monophone shared acoustic model generation unit, and an inter-word tri / bi / monophone shared acoustic model storage unit.

  Use bi / monophone between words Acoustic model storage unit uses bi / monophone between words, including any number of different phone models of triphone, biphone, and monophone respectively learned from triphone, biphone, and monophone learning data between words Memorize the acoustic model.

  The logical triphone generator generates a phoneme model name consisting of a triplet of a left phoneme, a central phoneme, and a right phoneme of a phonephone model of biphone and monophone existing in the inter-word bi / monophone use acoustic model (for example, in the case of biphone, a- by replacing the environment-independent symbol ("*" in the example above) included in the phoneme system with the left phoneme and / or right phoneme of monophone such as k + * Generate all triphone phoneme model names (for example, a-k + a, a-k + i, a-k + k, a-k + y, etc.) The phoneme model name that does not exist in the bi / monophone acoustic model is identified, and the model parameter of the phoneme model name of the phoneme model name of biphone or monophone that is the source of each identified phoneme model name is set to each specified phoneme It as model parameter of model name triphone phoneme model Re by copying, it generates a new phoneme models of triphone of each phoneme model name the specified.

  The inter-word tri / bi / monophone shared acoustic model storage unit captures and stores the inter-word bi / monophone use acoustic model and stores the triphone phoneme model newly generated by the logical triphone generation unit.

  According to the acoustic model creation device of the present invention, when an inter-word triphone use decoder and an inter-word bi / monophone use decoder are used, an acoustic model that can be shared by both decoders without compromising the speech recognition accuracy of each decoder is created. can do.

The figure which shows the apparatus structural example of the speech recognition process using the acoustic model creation apparatus of this invention. The figure which shows the function structural example of the acoustic model production apparatus of Example 1,3. The figure which shows the function structural example of the acoustic model production apparatus of Example 2, 5. FIG. FIG. 6 is a diagram illustrating a functional configuration example of an acoustic model creation device according to a fourth embodiment. FIG. 3 is a diagram illustrating a processing flow example of the acoustic model creation device according to the first embodiment. The figure which shows the example of a processing flow of the acoustic model production apparatus of Example 2. FIG. The figure which shows the example of a processing flow of the acoustic model production apparatus of Example 3. FIG. The figure which shows the example of a processing flow of the acoustic model production apparatus of Example 4. FIG. FIG. 10 is a diagram illustrating a processing flow example of the acoustic model creation device according to the fifth embodiment. The figure which shows the apparatus structural example of the conventional speech recognition process.

  FIG. 1 shows an apparatus configuration example for performing speech recognition processing using the acoustic model creation apparatus of the present invention. Although FIG. 1 shows the case where the acoustic model creation device 100 of the first embodiment is used, the acoustic model creation devices of other embodiments are also arranged at the same position.

  FIG. 2 shows a functional configuration example of the acoustic model creation apparatus 100 of the present invention, and FIG. 5 shows a processing flow example thereof. The acoustic model creation apparatus 100 includes an inter-word bi / monophone use acoustic model storage unit 110, a logical triphone generation unit 120, and an inter-word tri / bi / monophone shared acoustic model storage unit 130.

  The inter-word bi / monophone use acoustic model storage unit 110 includes an inter-word bi / monophone phoneme model including any number of triphone, biphone, and monophone phoneme models learned from the triphone, biphone, and monophone learning data. An acoustic model is stored in advance.

  First, the logical triphone generation unit 120 has a phoneme model name consisting of a triplet of a left phoneme, a central phoneme, and a right phoneme of a phonephone model of biphone and monophone existing in an interword bi / monophone use acoustic model (for example, in the case of biphone, Replace the environment-independent symbol ("*" in the above example) included in the phoneme system with the left phoneme and / or right phoneme of monophone such as a-k + * All the phone model names (for example, a-k + a, a-k + i, a-k + k, a-k + y, etc.) that can be generated are generated (S1). It should be noted that some phoneme combinations that can be generated do not appear in some languages (for example, k + k in Japanese), so such combinations may be excluded from generation targets). Subsequently, among the phoneme model names generated in S1, a phoneme model name that does not exist in the inter-word bi / monophone use acoustic model is specified (S2). For example, in the generation example given in S1, when a-k + a and a-k + i exist in the acoustic model, a-k + k and a-k + y are specified. Subsequently, model parameters of the phoneme model of the phoneme model name of the phone or monophone that is the generation source of each phoneme model name specified in S2 (the state transition probability and mixed normal distribution of each state of the phoneme model, and the mixed normal distribution) Each normal distribution, and the average vector and covariance matrix of each normal distribution) are copied as model parameters of the triphone phoneme model of each specified phoneme model name. A new phone model of triphone is generated (S3). For example, if the name of the phoneme model from which a-k + k is generated is a-k + *, copy the model parameters of the phone model of a-k + * as the model parameters of the phone model of a-k + k As a result, an a-k + k phoneme model is generated.

  The inter-word tri / bi / monophone shared acoustic model storage unit 130 takes in and stores the inter-word bi / monophone use acoustic model from the inter-word bi / monophone use acoustic model storage unit 110 and newly stores it in the logical triphone generation unit 120. The generated phone model of triphone is stored (S4).

  Thus, a triphone phoneme model that does not exist in the acoustic model is newly generated, and an inter-word tri / bi / monophone shared acoustic model to which the newly generated triphone phoneme model is added is used. Thus, since back-off does not occur, the inter-word triphone use decoder and the inter-word bi / monophone use decoder can share the acoustic model without impairing the speech recognition accuracy in each decoder.

  FIG. 3 shows a functional configuration example of the acoustic model creation apparatus 200 of the present invention, and FIG. 6 shows a processing flow example thereof. The acoustic model creation apparatus 200 includes an inter-word bi / monophone use acoustic model storage unit 110, a logical triphone generation unit 120, a logical biphone generation unit 220, and an inter-word tri / bi / monophone shared acoustic model storage unit 130. That is, the configuration is such that the logical biphone generation unit 220 is added to the acoustic model creation device 100 of the first embodiment. Thereby, since the phoneme model of the biphone which does not exist in the inter-word bi / monophone use acoustic model storage unit 110 can be generated and added to the acoustic model, it is possible to prevent the back-off of the biphone.

  First, the logical biphone generation unit 220 includes the left phoneme and the right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model existing in the inter-word bi / monophone use acoustic model. All phoneme model names that can be generated are generated by replacing any one of the environment-independent symbols with all phonemes in the phoneme system (S11). For example, in the case of * -a + * monophone, it is generated as k-a + *, y-a + *, * -a + i, * -a + n. It should be noted that some phoneme combinations that can be generated do not appear in some languages (for example, k + k in Japanese), so such combinations may be excluded from generation targets). Subsequently, among the phoneme model names generated in S11, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model are specified (S12). For example, when k-a + * and * -a + n exist in the acoustic model in the generation example given in S11, y-a + * and * -a + i are specified. Subsequently, by copying the model parameters of the phoneme model of the phonephone model name of the monophone that is the generation source of each phoneme model name specified in S12 as the model parameters of the phoneme model of the biphone of each specified phoneme model name Then, a new phoneme model of the biphone having the specified phoneme model name is generated (S13). For example, if the phoneme model name of the source of k-a + * is * -a + *, copy the model parameter of the phone model of * -a + * as the model parameter of the phoneme model of k-a + * , K-a + * phoneme model is generated.

  Then, the phoneme model of the biphone newly generated by the logical biphone generation unit 220 is further stored in the inter-word tri / bi / monophone shared acoustic model storage unit 130 (S4).

  FIG. 2 shows a functional configuration example of the acoustic model creation apparatus 300 of the present invention, and FIG. 7 shows a processing flow example thereof. The acoustic model creation apparatus 300 includes an inter-word bi / monophone use acoustic model storage unit 110, a logical triphone generation unit 320, and an inter-word tri / bi / monophone shared acoustic model storage unit 130. That is, the logical triphone generation unit 120 of the acoustic model creation device 100 according to the first embodiment is replaced with the logical triphone generation unit 320.

  First, the logical triphone generator 320 generates a phoneme model name consisting of a triplet of a left phoneme, a central phoneme, and a right phoneme of a phonephone model of biphone and monophone existing in an interword bi / monophone use acoustic model (for example, in the case of biphone). phone name of triphone that can be generated by replacing environment-independent symbols contained in left phoneme and / or right phoneme of monophone such as a-k + *, etc. with all phonemes in phoneme system (For example, a-k + a, a-k + i, a-k + k, a-k + y, etc.) are all generated (S21). It should be noted that some phoneme combinations that can be generated do not appear in some languages (for example, k + k in Japanese), so such combinations may be excluded from generation targets). Subsequently, among the phoneme model names generated in S21, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model are specified (S22). For example, in the generation example given in S21, when a-k + a and a-k + i exist in the acoustic model, a-k + k and a-k + y are specified. Subsequently, among the triphone phoneme model names specified in S22, those having the same phoneme model name of the source biphone or monophone, that is, those having the same back-off destination, the source biphone Alternatively, by expressing the environment-independent symbol part of the phone model name of monophone as a set of phonemes, the phoneme model set that integrally represents the phone model names of multiple triphones that share the phone model name of the source biphone or monophone A name is generated (S23). For example, when the phoneme model name of the generation source of a-k + k and a-k + y is a-k + *, a phoneme model set name of a-k + {*, k, y} is generated. Subsequently, as a model parameter common to the phoneme model of each phoneme model name constituting the phoneme model set name generated in S23, by copying the model parameter of the phoneme model name of the phonephone model name of the source biphone or monophone, A triphone phoneme model is newly generated in units of the phoneme model set name (S24). For example, as a model parameter common to each phoneme model name constituting the phoneme model set name a-k + {*, k, y}, ie, a-k + k, a-k + y, a The model parameters of the phone model of -k + * are copied, and a phoneme model is generated in units of phoneme model set names, that is, a set of a-k + {*, k, y}. The reason why the environment-independent symbol “*” is left in the phoneme model set name generated in S23 is that the phoneme model name of the corresponding generation source is back-off from the phoneme model set name in S24 (described above) In the example, a-k + *) is specified so that the model parameter of the specified phoneme model can be copied.

  The inter-word tri / bi / monophone shared acoustic model storage unit 130 takes in and stores the inter-word bi / monophone use acoustic model from the inter-word bi / monophone use acoustic model storage unit 110, and newly creates a logical triphone generation unit 320. The generated phone model of triphone is stored (S25). Note that the phoneme model set name of the triphone phoneme model newly generated by the logical triphone generation unit 320 is back-off as described above and the corresponding phoneme model name (a-k + * in the above example). The phoneme model having the phoneme model name is originally included in the inter-word bi / monophone use acoustic model imported from the inter-word bi / monophone use acoustic model storage unit 110. Therefore, when storing here, for example, the environment-independent symbol “*” of the phoneme model set name is converted to another symbol (for example, any) so as not to overlap (in the above example, a-k + {any, k, y}) and need to remember. In order to suppress an increase in size, another symbol used here (any in the above example) may be finally deleted.

  According to the acoustic model creation apparatus 300 described above, in addition to preventing the occurrence of back-off, model parameters can be set for each phoneme model set name for a phoneme model having a plurality of phoneme model names of a newly generated triphone. Since this is shared, an increase in the size of the acoustic model due to the addition of a new phoneme model name can be suppressed. Also, in the case of triphone, if model parameters are separated for each phoneme model name, the learning data amount of each phoneme model may be reduced, leading to deterioration of speech recognition accuracy. According to the present invention, Problems can also be avoided.

  FIG. 4 shows a functional configuration example of the acoustic model creation apparatus 400 of the present invention, and FIG. 8 shows a processing flow example thereof. The acoustic model creation apparatus 400 includes an inter-word bi / monophone use acoustic model storage unit 110, a logical triphone generation unit 420, and an inter-word tri / bi / monophone shared acoustic model storage unit 130. The acoustic model creation device 400 has a configuration in which the logical triphone generation unit 320 of the acoustic model creation device 300 is replaced with a logical triphone generation unit 420. Specifically, in the logical triphone generation unit 320, the processing is shared between the case of back-off from triphone to biphone and the case of back-off from triphone to monophone, whereas in the logical triphone generation unit 420, Those processes are performed individually.

  The logical triphone generation unit 420 includes logical bi / triphone generation means 421 and logical mono / triphone generation means 422.

  First, the logical bi / triphone generation means 421 is a phoneme model name (for example, a-k +) consisting of a triplet of a left phoneme, a central phoneme, and a right phoneme of a phoneme model of a biphone existing in an interword bi / monophone use acoustic model. *, Etc.) triphone phoneme model names that can be generated by replacing environment-independent symbols included in the left phoneme or right phoneme with all phonemes in the phoneme system (eg, a-k + a, a-k + i, a-k + k, a-k + y, etc.) are generated (S31). It should be noted that some phoneme combinations that can be generated do not appear in some languages (for example, k + k in Japanese), so such combinations may be excluded from generation targets). Subsequently, among the phoneme model names generated in S31, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model are specified (S32). For example, in the generation example given in S31, when a-k + a and a-k + i exist in the acoustic model, a-k + k and a-k + y are specified. Subsequently, the phoneme model name of the source biphone in the triphone phoneme model name specified in S32, that is, the phoneme model name of the source biphone is the same for the phonephone model name of the source biphone. By expressing the environment-independent symbol part of the model name as a phoneme set, a phoneme model set name that integrally represents the phoneme model names of a plurality of triphones that share the phoneme model name of the source biphone is generated (S33). ). For example, when the phoneme model name of the generation source of a-k + k and a-k + y is a-k + *, a phoneme model set name of a-k + {*, k, y} is generated. Subsequently, by copying the phoneme model model parameter of the phoneme model name of the source biphone as a model parameter common to the phoneme model of each phoneme model name constituting the phoneme model set name generated in S33, the phoneme model name is copied. A new phone model of triphone is generated in units of model set names (S34). For example, as a model parameter common to each phoneme model name constituting the phoneme model set name a-k + {*, k, y}, ie, a-k + k, a-k + y, a The model parameter of the phone model of -k + * is copied, and a phoneme model is generated in units of phoneme model set names, that is, a-k + {*, k, y}. The reason why the environment-independent symbol “*” is left in the phoneme model set name generated in S33 is that the phoneme model name of the corresponding generation source is back-off from the phoneme model set name in S34 (described above) In the example, a-k + *) is specified so that the model parameter of the specified phoneme model can be copied.

First, the logical mono / triphone generation means 422 is a phoneme model name (for example, * -k + *, Etc.) can be generated by replacing environment-independent symbols contained in the left and right phonemes with all phonemes in the phoneme system (eg, a-k + a, a-k + i, a-k + k, a-k + y, etc.) are generated (S35). It should be noted that some phoneme combinations that can be generated do not appear in some languages (for example, k + k in Japanese), so such combinations may be excluded from generation targets). Then, among the phoneme model name generated in S35, it does not exist in the inter-word bi / monophone using acoustic models and identifies the phoneme model name of triphone not specified in the logical bi / triphone raw formed means (S36). For example, in the generation example given in S35, when a-k + a and a-k + i exist in the acoustic model, a-k + k and a-k + y are specified. Subsequently, for the phoneme model name of the triphone identified in S36 that has the same phoneme model name of the source monophone, the left phoneme is assigned to the environment-independent symbol part of the phoneme model name of the source monophone. Is fixed to a certain phoneme, the right phoneme is expressed as a set of phonemes, and a phoneme model set name that integrally represents phoneme model names of multiple triphones that share the same phoneme model name of the source monophone is generated. The right phoneme is fixed to a certain phoneme, and the left phoneme is expressed as a set of phonemes to generate a phoneme model set name that integrally represents the phoneme model names of multiple triphones with the same monophone phoneme model name. If the number of phoneme model set names to be generated is smaller, the one with the smaller number is selected (S37). For example, the phoneme model name from which a triphone phoneme model name such as a-k + k, a-k + y, i-k + k, i-k + y, ... is generated is * -k + * In this case, a-k + {*, k, y}, i-k + {*, k, y}, ... is generated, with the left phoneme fixed and the right phoneme represented as a phoneme set as the phoneme model set name And the case of generating {*, a, i} -k + k, {*, a, i} -k + y, ... with the right phoneme fixed and the left phoneme represented as a set of phonemes Thus, the one with the smaller number of phoneme model set names to be generated is selected. The number of phoneme model set names varies depending on whether the left phoneme or right phoneme is fixed because there are phoneme combinations that do not appear depending on the language. Subsequently, by copying the model parameter of the phoneme model name of the phonephone model name of the original monophone as a model parameter common to the phoneme model name of the triphone of each phoneme model name constituting the phoneme model set name selected in S37, A triphone phoneme model is newly generated in units of the phoneme model set name (S38). For example, as a model parameter common to each phoneme model name constituting the phoneme model set name a-k + {*, k, y}, that is, a-k + k, a-k + y, The model parameter of the phone model of -k + * is copied, and a phoneme model is generated in units of phoneme model set names, that is, a-k + {*, k, y}. The reason why the environment-independent symbol “*” is left in the phoneme model set name generated in S37 is that the phoneme model name of the corresponding generation source is back-off from the phoneme model set name in S38 (described above) In the example, * -k + *) is specified so that the model parameter of the specified phoneme model can be copied.

  The inter-word tri / bi / monophone shared acoustic model storage unit 130 takes in and stores the inter-word bi / monophone use acoustic model from the inter-word bi / monophone use acoustic model storage unit 110, and logical bi / triphone generation means. The phone model of the triphone newly generated by the 421 and the logical mono / triphone generating means 422 is stored (S25). Note that the phoneme model set name (a-k + {*, k, y} in the above example) of the triphone phoneme model newly generated by the logical bi / triphone generation unit 421 and the logical mono / triphone generation unit 422 , Back-off and the corresponding phoneme model name (a-k + * in the above example) is included. The phoneme models with these phoneme model names are stored in the inter-word bi / monophone acoustic model memory. Originally included in the inter-word bi / monophone use acoustic model imported from the unit 110. Therefore, when storing here, for example, the environment-independent symbol “*” of the phoneme model set name is converted to another symbol (for example, any) so as not to overlap (in the above example, a-k + {any, k, y}) and need to remember. In order to suppress an increase in size, another symbol used here (any in the above example) may be finally deleted.

  According to the acoustic model creation apparatus 400 described above, in addition to preventing the occurrence of back-off, model parameters can be set for each phoneme model set name for a phoneme model with a plurality of phoneme model names of a newly generated triphone. Since this is shared, an increase in the size of the acoustic model due to the addition of a new phoneme model name can be suppressed. In particular, when the phoneme model name of the corresponding generation source is monophone after back-off, the number of phoneme model names newly generated is small when the left phoneme is fixed and when the right phoneme is fixed. Therefore, an increase in the size of the acoustic model can be further suppressed as compared with the third embodiment. Further, in the case of triphone, separating model parameters for each phoneme model name may reduce the amount of learning data for each phoneme model and cause deterioration in speech recognition accuracy. Can also be avoided.

  FIG. 3 shows a functional configuration example of the acoustic model creation apparatus 500 of the present invention, and FIG. 9 shows a processing flow example thereof. The acoustic model creation apparatus 500 includes an inter-word bi / monophone use acoustic model storage unit 110, a logical triphone generation unit 320, a logical biphone generation unit 520, and an inter-word tri / bi / monophone shared acoustic model storage unit 130. That is, this is a configuration in which the logical biphone generation unit 520 is added to the acoustic model creation device 300 of the third embodiment. Thereby, since the phoneme model of the biphone which does not exist in the inter-word bi / monophone use acoustic model storage unit 110 can be generated and added to the acoustic model, it is possible to prevent the back-off of the biphone. In addition, although the structure which added the logic biphone production | generation part 520 to the acoustic model creation apparatus 300 is demonstrated here, it can apply similarly to the acoustic model creation apparatus 400 of Example 4. FIG.

  First, the logical biphone generation unit 520 includes the left phoneme and the right phoneme of the phoneme model name composed of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model in the inter-word bi / monophone use acoustic model. All the phoneme model names that can be generated are generated by replacing any one of the environment-independent symbols with all phonemes in the phoneme system (S41). For example, in the case of * -a + * monophone, it is generated as k-a + *, y-a + *, * -a + i, * -a + n. It should be noted that some phoneme combinations that can be generated do not appear in some languages (for example, k + k in Japanese), so such combinations may be excluded from generation targets). Subsequently, among the phoneme model names generated in S41, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model are specified (S42). For example, in the generation example given in S41, when k-a + * and * -a + n exist in the acoustic model, y-a + * and * -a + i are specified. Subsequently, for the phoneme model name of the biphone identified in S42, the phoneme model name of the source monophone is the same, and either one of the environment-independent symbols of the phoneme model name of the source monophone By expressing the part as a set of phonemes, a phoneme model set name that integrally represents the phoneme model names of a plurality of biphones with the same phoneme model name of the source monophone is generated (S43). For example, if the phoneme model name from which a-k + * and i-k + * are generated is * -k + *, a phoneme model set name {*, a, i} -k + * is generated. Subsequently, as a model parameter common to the phoneme model of each phoneme model name constituting the phoneme model set name generated in S43, the phoneme model model parameter of the phoneme model name of the source monophone is copied to thereby obtain the phoneme model name. A new phoneme model of biphone is generated in units of model set names (S44). For example, as a model parameter common to each phoneme model name constituting the phoneme model set name {*, a, i} -k + *, that is, a-k + *, i-k + *, * -k + of the generation source The model parameters of the phoneme model of * are copied, and a phoneme model is generated with a phoneme model set name, that is, a set unit of {*, a, i} -k + *. The reason why the environment-independent symbol “*” is left in the phoneme model set name generated in S43 is that the phoneme model name of the corresponding generation source is back-off from the phoneme model set name in S44 (described above) In the example, * -k + *) is specified so that the model parameter of the specified phoneme model can be copied.

  Then, the inter-word tri / bi / monophone shared acoustic model storage unit 130 further stores the phoneme model of the biphone newly generated by the logical biphone generation unit 520 (S25).

  When an acoustic model that uses transition probabilities is used in a decoder that does not use transition probabilities, mismatch occurs between acoustic model learning and speech recognition, and recognition accuracy deteriorates. Therefore, prior to storing the triphone phoneme model in the inter-word tri / bi / monophone shared acoustic model storage unit 130 of the acoustic model creation apparatuses 100 to 500 of the first to fifth embodiments, the triphone phoneme model uses the transition probability. In the case of a phoneme model, it is conceivable to perform a conversion process by providing a transition probability conversion unit 610 that converts a phoneme model that does not use a transition probability or a phoneme model that has a predetermined transition probability value (for example, 0.5). (S51). As a result, when the constructed acoustic model uses transition probabilities, both decoders using and not using transition probabilities can be used while maintaining speech recognition accuracy.

  The newly generated unlearned triphone and biphone are not necessarily suitable for the model parameters, but high accuracy can be realized by performing additional learning.

  Therefore, prior to storing the phoneme model in the inter-word tri / bi / monophone shared acoustic model storage unit 130 of the acoustic model creation apparatuses 100 to 500 of the first to sixth embodiments, additional learning is performed on the newly generated phoneme model. It is conceivable to perform an additional learning process by providing an additional learning unit 710 to perform (S61).

  It should be noted that, by implementing the present invention, the phoneme sharing relationship is generated and changed, so that the alignment for associating the learning data with the phoneme model changes. It is desirable to additionally learn not only the phoneme model that has been set, but also the existing phoneme model. When the transition probability conversion unit 610 is provided, the transition probability may be ignored only during additional learning.

  When the acoustic model creation apparatus of the present invention is realized by a computer, the processing contents of the functions of the apparatus and each part thereof are described by a program. The program is stored in, for example, a hard disk device, and necessary programs and data are read into a RAM (Random Access Memory) at the time of execution. The read program is executed by the CPU, whereby each processing content is realized on the computer. Note that at least a part of the processing content may be realized by hardware.

  Note that the program in the present invention includes information provided for processing by an electronic computer that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

Claims (15)

Inter-word bi / monophone use acoustic model that stores an inter-word bi / monophone use acoustic model, including any number of different phone models of triphone, biphone, monophone respectively learned by triphone, biphone, monophone learning data between words A model storage unit;
Environment independent of phoneme model name consisting of triplet of left phoneme, central phoneme, right phoneme of phoneme model of biphone and monophone existing in acoustic model using inter / biphone / monophone, included as left phoneme and / or right phoneme All the phoneme model names that can be generated by replacing symbols with all phonemes in the phoneme system are generated, and among the generated phoneme model names, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model are identified By copying the model parameters of the phone model of the phone model name of biphone or monophone that is the generation source of each specified phone model name as the model parameters of the phone model of the tri phone of each specified phone model name , A logical triphone generation unit for newly generating a triphone phoneme model of each identified phoneme model name,
The inter-bi bi / monophone use acoustic model is captured and stored, and the inter-tri / bi / monophone acoustic model storage unit stores the triphone phoneme model newly generated by the logical triphone generation unit;
An acoustic model creation device comprising: The acoustic model creation device according to claim 1,
Furthermore, the environment-independent symbols included as the left phoneme and right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model present in the inter-word bi / monophone use acoustic model All phoneme model names that can be generated by replacing either one with all phonemes in the phoneme system are generated, and among the generated phoneme model names, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model By copying the model parameters of the phone model of the phone model name of the monophone that is the source of each specified phone model name as the model parameters of the phone model of the phone model of each specified phone model name A logical biphone generation unit for newly generating a phoneme model of the biphone of each identified phoneme model name,
The inter-word tri / bi / monophone shared acoustic model storage unit further stores a phoneme model of a biphone newly generated by a logical biphone generation unit. Inter-word bi / monophone use acoustic model that stores an inter-word bi / monophone use acoustic model, including any number of different phone models of triphone, biphone, monophone respectively learned by triphone, biphone, monophone learning data between words A model storage unit;
Environment independent of phoneme model name consisting of triplet of left phoneme, central phoneme, right phoneme of phoneme model of biphone and monophone existing in acoustic model using inter / biphone / monophone, included as left phoneme and / or right phoneme All the phoneme model names that can be generated by replacing symbols with all phonemes in the phoneme system are generated, and among the generated phoneme model names, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model are identified If the phone model name of the source biphone or monophone is the same among the specified phone names of the triphone, the environment-independent symbol part of the phone model name of the source biphone or monophone is used as the phoneme. The phoneme model set name that integrally represents the phone model names of multiple triphones that share the same phoneme model name of the source biphone or monophone. The phoneme model is generated by copying the phoneme model model parameter of the phoneme model name of the source biphone or monophone as a model parameter common to the phoneme model of each phoneme model name constituting the phoneme model set name. A logical triphone generator that newly generates a phone model of triphone in units of set names;
The inter-bi bi / monophone use acoustic model is captured and stored, and the inter-tri / bi / monophone acoustic model storage unit stores the triphone phoneme model newly generated by the logical triphone generation unit;
An acoustic model creation device comprising: The acoustic model creation device according to claim 3,
The logical triphone generator is
The phoneme system includes environment-independent symbols included as the left phoneme or the right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the biphone phoneme model present in the inter-word bi / monophone use acoustic model All phoneme model names that can be generated by replacing all phonemes inside are generated, and among the generated phoneme model names, phoneme model names that do not exist in the inter-word bi / monophone acoustic model are specified and specified By expressing the environment-independent symbol part of the phoneme model name of the source biphone as a set of phonemes for the phoneme model name of the source biphone in the triphone phoneme model name Generate phoneme model set names that collectively represent the phoneme model names of multiple triphones that have the same phoneme model name of the source biphone, and generate each phoneme model set name As a model parameter common to the phoneme model of the elementary model name, by copying the model parameter of the phoneme model of the phonemic model name of the source biphone, a new phoneme model of triphone is generated in units of the phoneme model set name Logical bi / triphone generation means;
The phoneme system includes environment-independent symbols included as the left phoneme and right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model existing in the inter-word bi / monophone use acoustic model All phoneme model names of triphones that can be generated are generated by replacing all phonemes in the phoneme, and among the generated phoneme model names, the phonetic model name does not exist in the inter-word bi / monophone use acoustic model and the logical bi / triphone identify the phoneme model name of triphone not specified in generating means, about what phoneme model name of the originator of monophone in phoneme model name of the specified triphone are common, of the generator of monophone For the environment-independent symbol part of the phoneme model name, the left phoneme is fixed to a certain phoneme and the right phoneme is expressed as a set of phonemes, and multiple triphones with the same phoneme model name of the source monophone When the phoneme model set name that represents the phoneme model name is generated, and the right phoneme is fixed to a certain phoneme and the left phoneme is expressed as a set of phonemes, the phoneme model name of the source monophone is the same Select the phoneme model set name with the smaller number of phoneme model set names when the phoneme model set names that collectively represent the phoneme model names of multiple triphones are generated, and select the phoneme model set names that make up the selected phoneme model set name As a model parameter common to the phone name model of the triphone of the model name, a phoneme model of the triphone phoneme model is generated by copying the model parameter of the phone model of the phone model name of the source monophone. Logical mono / triphone generation means to
An acoustic model creation device comprising: In the acoustic model creation device according to claim 3 or 4,
Furthermore, the environment-independent symbols included as the left phoneme and right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model present in the inter-word bi / monophone use acoustic model All phoneme model names that can be generated by replacing either one with all phonemes in the phoneme system are generated, and among the generated phoneme model names, phoneme model names that do not exist in the inter-word bi / monophone use acoustic model The phoneme model name of the source monophone in the specified phoneme model name of the specified biphone is the same as the environment-independent symbol of either one of the phoneme model name of the source monophone By expressing the part as a set of phonemes, generate a phoneme model set name that integrally represents the phoneme model names of multiple biphones that have the same phoneme model name of the source monophone By copying the phoneme model model parameter of the phone model name of the original monophone as a model parameter common to the phoneme model of each phoneme model name constituting the generated phoneme model set name, It has a logical biphone generation unit that newly generates phoneme models of biphone in units,
The inter-word tri / bi / monophone shared acoustic model storage unit further stores a phoneme model of a biphone newly generated by a logical biphone generation unit. In the acoustic model creation device according to any one of claims 1 to 5 ,
Before storing the triphone phoneme model in the tri / bi / monophon e shared acoustic model storage unit between words, if the phone phone model of the triphone is a phoneme model that uses a transition probability, a phoneme model that does not use the transition probability or a predetermined one An acoustic model creation apparatus, further comprising a transition probability conversion unit that converts a phoneme model having a transition probability value of. In the acoustic model creation device according to any one of claims 1 to 6 ,
An acoustic model creation device further comprising an additional learning unit that performs additional learning on a newly generated phoneme model prior to storing the phoneme model in the inter-word tri / bi / monophon e shared acoustic model storage unit . Inter-word bi / monophone use acoustic model that stores an inter-word bi / monophone use acoustic model, including any number of different phone models of triphone, biphone, monophone respectively learned by triphone, biphone, monophone learning data between words Using the model storage unit,
Environment independent of phoneme model name consisting of triplet of left phoneme, central phoneme, right phoneme of phoneme model of biphone and monophone existing in acoustic model using inter / biphone / monophone, included as left phoneme and / or right phoneme A triphone all-phoneme model name generation step for generating all phone names of triphones that can be generated by replacing symbols with all phonemes in the phoneme system;
Among the generated phoneme model names, a triphone phoneme model name specifying step for specifying a phoneme model name that does not exist in the inter-word bi / monophone use acoustic model,
The phoneme model model parameter of the phonephone model name of biphone or monophone that is the source of each identified phoneme model name is copied as the model parameter of the triphone phoneme model of each specified phoneme model name A triphone phoneme model generation step for newly generating a triphone phoneme model of each phoneme model name,
The inter-word tri / bi / monophone shared acoustic model storage unit captures and stores the inter-word bi / monophone use acoustic model and stores the triphone phoneme model newly generated by the logical triphone generation unit. / bi / monophone shared acoustic model storage step,
Acoustic model creation method to execute. The acoustic model creation method according to claim 8,
Furthermore, the environment-independent symbols included as the left phoneme and right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model present in the inter-word bi / monophone use acoustic model A biphone full phoneme model name generation step for generating all phoneme model names of biphone that can be generated by replacing either one with all phonemes in the phoneme system;
Among the generated phoneme model names, a biphone phoneme model name identifying step for identifying a phoneme model name that does not exist in the inter-word bi / monophone use acoustic model,
Each identified phoneme model name is generated by copying the phoneme model model parameters of the phonephone model name of the monophone that is the source of each identified phoneme model name as model parameters of the phonephone model of the identified phoneme model name. A biphone phoneme model generation step for newly generating a phoneme model of the phoneme model name biphone;
Run
The inter-word tri / bi / monophone shared acoustic model storing step further stores a phonemic model of biphone newly generated in the biphone phoneme model generating step. Inter-word bi / monophone use acoustic model that stores an inter-word bi / monophone use acoustic model, including any number of different phone models of triphone, biphone, monophone respectively learned by triphone, biphone, monophone learning data between words Using the model storage unit,
Environment independent of phoneme model name consisting of triplet of left phoneme, central phoneme, right phoneme of phoneme model of biphone and monophone existing in acoustic model using inter / biphone / monophone, included as left phoneme and / or right phoneme A triphone all-phoneme model name generation step for generating all phone names of triphones that can be generated by replacing symbols with all phonemes in the phoneme system;
Among the generated phoneme model names, a triphone phoneme model name specifying step for specifying a phoneme model name that does not exist in the inter-word bi / monophone use acoustic model,
Among the phoneme model names of the specified triphone that have the same phoneme model name of the source biphone or monophone, use the environment-independent symbol part of the phoneme model name of the source biphone or monophone as the set of phonemes A triphone phoneme model set name generation step for generating a phoneme model set name that integrally represents the phoneme model names of a plurality of triphones that share the phoneme model name of the source biphone or monophone by expressing,
As a model parameter common to the phoneme model of each phoneme model name constituting the generated phoneme model set name, copy the phoneme model model parameter of the phonephone model name of the source biphone or monophone, and thereby name the phoneme model set A triphone phoneme model generation step for newly generating a triphone phoneme model in units of
The inter-word tri / bi / monophone shared acoustic model storage unit captures and stores the inter-word bi / monophone use acoustic model, and stores the triphone phoneme model newly generated in the triphone phoneme model generation step. Tri / bi / monophone shared acoustic model storage step,
An acoustic model creation method comprising: The acoustic model creation method according to claim 10 ,
The phoneme system includes environment-independent symbols included as the left phoneme or the right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the biphone phoneme model present in the inter-word bi / monophone use acoustic model A first triphone all-phoneme model name generation substep for generating all triphone phoneme model names that can be generated by replacing all the phonemes in
A first triphone phoneme model name identification substep for identifying a phoneme model name that does not exist in the inter-word bi / monophone use acoustic model among the phoneme model names generated in the first triphone all-phoneme model name generation substep;
The phoneme model name of triphone identified in the first triphone phoneme model name identification sub- step, where the phoneme model name of the source biphone is common, the environment-independent symbol part of the phoneme model name of the source biphone A first triphone phoneme model set name generation sub-step for generating a phoneme model set name that integrally represents a plurality of triphone phoneme model names having the same phoneme model name of the source biphone ,
The model parameter of the phoneme model of the phoneme model name of the source biphone is copied as a model parameter common to the phoneme model of each phoneme model name constituting the phoneme model set name generated in the first triphone phoneme model set name generation substep A first triphone phoneme model generation sub-step for newly generating a triphone phoneme model in units of the phoneme model set name;
Environment-independent symbols included as the left phoneme and right phoneme of the phoneme model name composed of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model existing in the inter-word bi / monophone use acoustic model, A second triphone all-phoneme model name generation substep for generating all triphone phoneme model names that can be generated by replacing all phonemes in the system;
Of the phoneme model names generated in the second triphone all-phoneme model name generation substep, triphones that do not exist in the inter-word bi / monophone use acoustic model and are not specified in the first triphone phoneme model name specification substep A second triphone phoneme model name identification substep for identifying a phoneme model name;
For the phone name of the triphone phoneme model specified in the second triphone phoneme model name specifying sub-step, the phoneme model name of the source monophone is the same, and the environment-independent symbol part of the phone phone model name of the source phone The left phoneme is fixed to a certain phoneme, the right phoneme is expressed as a set of phonemes, and the phoneme model set name that integrally represents the phoneme model names of multiple triphones that share the same phoneme model name of the source monophone. A phoneme model in which the right phoneme is fixed to a certain phoneme and the left phoneme is expressed as a set of phonemes, and the phoneme model names of multiple triphones with the same name are generated A second triphone phoneme model set name generation sub-step for selecting a set with a smaller number of generated phoneme model set names;
As a model parameter common to the phone model of the triphone phoneme model name constituting the phoneme model set name selected in the second triphone phoneme model set name generation sub-step, the model parameter of the phone model of the phone model name of the source monophone is used. A second triphone phoneme model generation sub- step for newly generating a triphone phoneme model in units of the phoneme model set name by copying;
The acoustic model creation method characterized by performing. The acoustic model creation method according to claim 10 or 11,
Furthermore, the environment-independent symbols included as the left phoneme and right phoneme of the phoneme model name consisting of the left phoneme, the central phoneme, and the right phoneme of the monophone phoneme model present in the inter-word bi / monophone use acoustic model A biphone full phoneme model name generation step for generating all phoneme model names of biphone that can be generated by replacing either one with all phonemes in the phoneme system;
Among the phoneme model names generated in the biphone all phoneme model name generation step, the biphone phoneme model name specifying step for specifying a phoneme model name that does not exist in the inter-word bi / monophone use acoustic model,
Among the phoneme model names specified in the biphone phoneme model name specifying step that have the same phoneme model name of the source monophone, the environment of either one of the phoneme model names of the source monophone Biphone phoneme model set name generation step for generating a phoneme model set name that integrally represents the phoneme model names of multiple biphones that share the phoneme model name of the source monophone by expressing the dependency symbol part as a set of phonemes When,
By copying the phoneme model model parameter of the phone model name of the original monophone as a model parameter common to the phoneme model name of each phoneme model name constituting the phoneme model set name generated in the biphone phoneme model name generation step , A biphone phoneme model generation step for newly generating a phoneme model of biphone in the unit of the phoneme model set name,
Run
The inter-word tri / bi / monophone shared acoustic model storing step further stores a phonemic model of biphone newly generated in the biphone phoneme model generating step. The acoustic model creation method according to any one of claims 8 to 12 ,
The inter-word tri / bi / monophone acoustic model storage step is a phoneme model in which the triphone phoneme model uses the transition probability prior to storing the triphone phoneme model in the inter-word tri / bi / monophone acoustic model storage unit. In some cases, the acoustic model creation method further includes a transition probability conversion step of converting into a phoneme model that does not use a transition probability or a phoneme model having a predetermined transition probability value. The acoustic model creation method according to any one of claims 8 to 13 ,
Additional learning to perform additional learning on the newly generated phoneme model prior to storing the phoneme model in the inter-word tri / bi / monophone shared acoustic model storage unit in the inter-word tri / bi / monophone shared acoustic model storage step A method for creating an acoustic model, further comprising executing a step. A program for causing a computer to execute the acoustic model creation method according to any one of claims 8 to 14 .

Images