JP4964194B2 - Speech recognition model creation device and method thereof, speech recognition device and method thereof, program and recording medium thereof - Google Patents

Description

  The present invention relates to a speech recognition model creation apparatus and method for efficiently learning a speech recognition model, a speech recognition apparatus and speech recognition method using the method, a program, and a recording medium.

  The speech recognition apparatus calculates a likelihood between a sequence of acoustic feature vectors obtained by analyzing an input speech signal and an acoustic model obtained by modeling speech, and recognizes between words to be recognized and words A word string having the highest likelihood is output as a recognition result in terms of linguistic constraints such as ease of connection and a language model representing a rule. The voice targeted for voice recognition generally changes its characteristics greatly depending on the external environment such as various speakers, languages and noise. In order to recognize speech having such various features, speech recognition methods that perform speech recognition using a plurality of speech recognition models have been widely studied.

  For example, Non-Patent Document 1 is intended for speech recognition in which English and German are mixed as an acoustic environment, and by implementing two types of speech recognition models for both languages, robust speech recognition is realized regarding language switching. It is shown. In Non-Patent Document 2, there is an example in which robust voice recognition with respect to speaker switching is realized by preparing a plurality of voice recognition models for voice recognition of a mixture of plural speakers in a conference in which many people participate. It is shown. In the example of Non-Patent Document 2, the performance improvement of the speech recognition apparatus is realized by performing adaptive learning for each speech recognition model of each speaker. Adaptive learning refers to adapting a limited speech recognition model recorded in a speech recognition device to acoustic features that vary depending on the speaker and the environment in the scene where it is actually used.

  An example of the functional configuration of a speech recognition apparatus 900 prepared with a plurality of conventional acoustic models will be briefly described with reference to FIG. The speech recognition apparatus 900 includes a speech recognition model 90, an A / D conversion unit 91, a feature amount extraction unit 92, a speech recognition unit 93, and an adaptive learning unit 94.

  The speech recognition model 90 is a speech recognition model corresponding to, for example, a plurality of languages and a plurality of speakers. For example, a first speech recognition model 901 for a certain speaker is composed of a first acoustic model memory 901a, a first language model memory 901b, and a first utterance dictionary model memory 901c. Similarly, the second speech recognition model 902 for other speakers includes a second acoustic model memory 902a, a second language model memory 902b, and a second utterance dictionary model memory 902c.

The A / D conversion unit 91 converts the sound of the input analog signal into a discrete digital signal with a sampling frequency of 16 kHz, for example. The feature amount extraction unit 92 extracts 320 feature amount vectors from the sound signal for each frame, for example, 320 discrete speech signals are regarded as one frame (20 ms). The feature vector is extracted by, for example, Mel frequency cepstrum coefficient (MFCC) analysis. The voice recognition unit 93 includes a score calculation unit 931 and a word string search unit 932. The score calculation unit 931 calculates a score for the feature vector by using the feature vector, the language model from the speech recognition model 901, and the acoustic model as inputs. The word string search unit 932 searches the utterance dictionary model memory 901c for a word string having the maximum score, and outputs it as a recognition result. The adaptive learning unit 94 performs an adaptive process for each of the first speech recognition model 901 and the second speech recognition model 902 using the word sequence output from the word sequence search unit 932 as a teacher signal.
Z.Wamg, U.Topkara, T.Schultz, and A.Waibel.Towards universal speech recognition.In Proc.ICMI2002,2002. Ryuta Takuma, Kouji Iwano, Sadahiro Furui “Study on Parallel Processing Conference Speech Recognition System Using Sequential Speaker Adaptation” Proceedings of the Spring Acoustical Society, p105-106, 2002.

  In adaptive learning of multiple speech recognition models by the conventional method, adaptive learning is performed independently for each speech recognition model, so adaptive learning data is distributed to multiple models, and the amount of allocated data is a single model. Compared to the adaptive learning of. For this reason, there is a problem that the effect of adaptive learning becomes limited due to a decrease in the data amount.

  The present invention has been made in view of these points, and a speech recognition model creation apparatus and method capable of efficiently performing adaptive learning of a plurality of speech recognition models, a speech recognition apparatus using the method, and speech recognition. It is an object to provide a method, a program, and a recording medium thereof.

The speech recognition model creation device according to the present invention includes an initial value speech recognition model recording unit, a likelihood calculation unit, a model update unit, and an updated speech recognition model recording unit. The initial value speech recognition model recording unit is an initial value speech recognition model that is a vector obtained by concatenating vectors representing parameters of a plurality of speech recognition models, and the plurality of speech recognition models are respectively connected to a plurality of sound sources. An initial value speech recognition model , which is a corresponding speech recognition model, is recorded. The likelihood calculation unit, feature amounts of the plurality of sets and the audio signal of the state sequence is the result of voice recognition a voice signal voice recognition network based on union operation state transition probability respectively corresponding to each speech recognition model Using the vector as an input , the likelihood of each state for each frame is calculated. Model updating unit as inputs the likelihood and the feature amount vector, to generate the updates to update the speech recognition model initial values speech recognition models. The updated speech recognition model recording unit records the updated speech recognition model.

  The speech recognition model creation apparatus according to the present invention handles an initial value speech recognition model including a plurality of speech recognition models as one vector. Then, the initial value speech recognition model is updated using a speech recognition result that is speech-recognized based on a state probability transition composed of a combination of a plurality of speech recognition models. That is, since a plurality of speech recognition models can be learned together, a sufficient adaptive learning effect can be obtained even with a small amount of speech data.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

[Basic idea of the present invention]
The basic idea of the voice recognition model creation method of the present invention will be described. Probabilistic speech recognition methods that are widely used nowadays use a probabilistic model to express the speech recognition process as the appearance probability (likelihood function) of speech data and words (or phonemes, HMM (Hidden Markov Model)), and the posterior probability. This is a method for estimating parameters for speech recognition using probability statistical evaluation criteria such as maximization and likelihood maximization. The speech recognition model creation method of the present invention is the same in the portion using this probability statistical evaluation criterion.

  The present invention is different from the conventional method in that an initial value speech recognition model including a plurality of speech recognition models is treated as one vector, and the initial value speech recognition model is treated as a state probability transition composed of a combination of a plurality of speech recognition models. This is a part for adaptive learning using a speech recognition result recognized based on speech. Here, the speech recognition result corresponds to a speech symbol sequence such as a word sequence, a phoneme sequence, or an HMM state sequence. These are collectively called a state sequence. In addition, the state sequence is expressed not only as a single series, but also as a set from a set of state sequences of n high-order scores (n-best), a sub-network of a speech recognition network such as lattice. These are collectively referred to as a set of state sequences.

The time series set O = {O _{t = 1} , O _{t = 2} ,...] Of the feature quantity vectors of the acoustic feature quantity output by the feature quantity extraction unit 92 is a subset O of the time series set output by one sound source A. _{e = 1} = {O _{e = 1, t = 1} , O _{e = 1, t = 2} ,...} and a subset O _{e = 2} = {O _{e = 2 , T = 1} , O _{e = 2, t = 2} ,. Here, a speech recognition model corresponding to the sound source A is e = 1, and a speech recognition model corresponding to the sound source B is e = 2. Here, two sound sources are targeted for ease of explanation, but the same applies to the case of three or more sound sources. The time series set of feature vectors output by the sound source A is O _{e = 1} , the speech recognition model is Θ _{e = 1} , the hidden variables are Z _{e = 1} = {Z _{e = 1, t = 1} , Z _{e = 1, Let t = 2} ,…}. Here, a hidden variable is a variable that cannot be observed to which object it belongs. In the case of a speech recognition model using HMM, the hidden variable Z _{e = 1} represents the ID of the HMM state at each frame time. At this time, the likelihood function of complete data can be expressed by equation (1).

Similarly for the speech recognition model e = 2, the likelihood function of complete data can be expressed by equation (2).

Thus, if the subsets O _{e = 1} and O _{e = 2} of the time series sets of each model are given in advance, the likelihood function can be given independently. However, in general, it is not known whether the sound data to be recognized is sound source A or sound source B. Therefore, in the present invention, a hidden variable U _t = {Z _{e = 1, t} , Z _{e = 2, t} ,} indicating whether the sound of the sound source A or the sound source B appears at each frame time t is newly set. Introduce. As a result, the entire hidden variable is composed of Z _{e = 1} , Z _{e = 2} , and U = {U _{t = 1} , U _{t = 2} ,.

ここで○は合成演算を表し、＊はネットワークのループを表現する。 As a possible value of Z, a score that takes into consideration pronunciation rules (pronunciation dictionary model) and ease of connection of words (language model), rather than an arbitrary HMM state sequence actually appearing at each time A state series on the voice recognition network with a mark appears. A speech recognition network (time series of state probability transitions) generally used in speech recognition is composed of a combination of three networks: HMM (H), dictionary (L), and grammar (G). The speech recognition network N is expressed by Expression (4) obtained by performing a synthesis operation on these networks.

Here, ○ represents a composition operation, and * represents a network loop.

When considering the conversation environment of the same language and two speakers, it is only necessary to prepare a time series of state probability transitions having the same dictionary and grammar model but different only in the HMM network. The speech recognition network N in this case can be constructed by combining two speech recognition networks as shown in Expression (5), assuming that a transition occurs between sentences.

Here, U (+) is a binary operation for a network in which two networks are matched with the start and end using a union operation. (+) Represents a union operation. (+) Is correct in the expression. FIG. 1 conceptually shows the union operation. N ₁ is a network having 14 states and 27 arcs, and N ₂ is a network having 8 states and 8 arcs. When the unions of the networks N ₁ and N ₂ are performed, the two networks can be described in parallel by matching the start and end points. The calculation of equation (5) means giving two choices of speech recognition network (H _{e = 1} · L · G) and (H _{e = 2} · L · G). As to which network is selected, a network having a high score such as a posteriori probability value or likelihood value in the search process is selected. In this way, by performing a search on the synthesized network expressed by Expression (5), a speech recognition result and a set of state sequences for learning to be described later when a plurality of sound sources are mixed are obtained. Can do.

The present invention includes a speech recognition network database that records a speech recognition network composed of a combination of a plurality of speech recognition models as shown in equation (5). In addition, an initial value speech recognition model μ ⁰ shown in Expression (6) is provided as an initial value of the speech recognition model. The average vector μ of the speech recognition model updated by adaptive learning is also handled as the same vector as that in Expression (6).

Equation (6) shows only the average vector, but the variance matrix Σ, the mixture weight w, and the state transition probability a are similarly collected as vectors, whereby a plurality of speech recognition models are handled as one vector. When the initial value speech recognition model Θ in which the plurality of speech recognition models are handled as one vector and the hidden variable Z in Equation (3) are used, the likelihood function can be expressed by Equation (7).

  As described above, in the present invention, using the synthesized network Z and the model parameter Θ, speech recognition in a plurality of environments is performed with the same likelihood as the single speech recognition model expressed by the equations (1) and (2). It can be realized using a function. That is, by using Expression (7), it is possible to realize speech recognition in a plurality of environments without changing the speech recognition decoder (speech recognition unit).

The speech recognition model creation method of the present invention is based on the likelihood function shown in Equation (8) from the likelihood function of Equation (7) and the initial value speech recognition model Θ in which a plurality of speech recognition models are treated as one vector. An updated speech recognition model Θ￣ used in the above is generated.

  The initial value speech recognition model Θ and the updated speech recognition model Θ￣ are expressed parametrically using the function F (•) and are related by the relation parameter φ. As described above, since the speech recognition model creation method of the present invention can learn a plurality of speech recognition models as one vector at a time, a sufficient adaptive learning effect can be obtained even with a small amount of speech data.

  FIG. 2 shows a functional configuration example of the speech recognition model creation device 100 of the present invention and the speech recognition device 200 having the same as a component. FIG. 3 shows an operation flow of the speech recognition model creation apparatus 100. The operation of the speech recognition model creation device 100 will be described with reference to FIGS.

  The speech recognition model creation apparatus 100 includes an initial value speech recognition model recording unit 10, a model update unit 12, a likelihood calculation unit 13, an updated speech recognition model recording unit 14, and a control unit 16. The speech recognition model creation device 100 and the speech recognition device 200 are realized by a predetermined program being read into a computer including, for example, a ROM, a RAM, and a CPU, and the CPU executing the program.

  The initial value speech recognition model recording unit 10 records an initial value speech recognition model including a plurality of speech recognition models. The likelihood calculating unit 13 calculates the likelihood of each state for each frame by using as input a set of state sequences that are speech-recognized based on state probability transitions composed of a combination of a plurality of speech recognition models (step S13).

Here, the relationship between the frame, each state, the Gaussian distribution, and the state probability transition will be described. The phoneme model constituting the speech recognition model is constructed according to the state shown in FIG. Each state i is expressed as a mixed normal distribution M _i . The mixed normal distribution M _i is composed of, for example, three normal distributions, N (μ _i1 , Σ _i1 ), N (μ _i2 , Σ _i2 ), and N (μ _i3 , Σ _i3 ).

The phoneme model is constructed by a probability chain of several to about a dozen of states i. FIG. 5 shows a conceptual diagram of a phoneme model composed of three states as an example. The example shown in FIG.
This is called an “ight type HMM”, which is an array of three states i ₁ (first state), i ₂ (second state), and i ₃ (third state), as a state probability chain (state transition). Consists of self transitions a ₁₁ , a ₂₂ , a ₃₃ and a ₁₂ , a ₂₃ , a ₃₄ to the next state. FIG. 6 shows a time-series relationship between the state i and the frame t. The horizontal axis represents the passage of time and is represented by a frame number. The vertical axis represents the state i of each frame. Each state i has a mixed normal distribution as shown in FIG. ● is the maximum likelihood state in which the output probability score is maximized within each frame. The maximum likelihood state sequence is the maximum likelihood state sequence arranged in time series. This maximum likelihood state sequence is output as a speech recognition result.

The likelihood calculating unit 13 obtains the likelihood p (O, Z _t = i | Θ ^) of each state i using, for example, a forward / backward algorithm. The likelihood p (O, Z _t = i | Θ ^) of each state i can be calculated by the equation (9) using the forward coefficient α and the backward coefficient β. The likelihood p and the feature quantity vector O may be recorded by sequentially storing those values obtained in advance by the speech recognition apparatus without being calculated again by the likelihood calculating unit 13.

The forward coefficient α and the backward coefficient β are calculated by Expressions (10) and (11) by iterative calculation in the maximum likelihood estimation method (EM algorithm).

Here, k is an index of the Gaussian distribution constituting the state i. a _ij represents a state transition probability when the state i transitions from i to j, w _jk represents a mixture weight factor for the Gaussian distribution k in the state j, N represents a Gaussian distribution of the mean vector μ _jk and the covariance matrix Σ _jk . ^ In Equations (10) and (11) indicates that each parameter described above is a value estimated in a step before the iterative calculation in the expected value maximization method.

  The model updating unit 12 generates an updated speech recognition model in which the initial value speech recognition model is updated as one vector using the likelihood as an input (step S12). The updated speech recognition model recording unit 14 records the updated speech recognition model (step S14). The model updating unit 12 and the updated speech recognition model recording unit 14 continue to operate until the control unit 16 issues a signal instructing the end of the operation (N in step S16).

  As described above, the model update unit 12 performs adaptive learning by treating the initial value speech recognition model including a plurality of speech recognition models as one vector, and thus sufficient adaptive learning effects can be obtained even with a small amount of speech data. . FIG. 7 shows a detailed functional configuration example of the model update unit 12 of the speech recognition model creation apparatus 100 and will be described in more detail.

The model update unit 12 includes a posterior probability calculation unit 121, a relationship parameter generation unit 122, and an update model generation unit 123. The posterior probability calculation unit 121 obtains the posterior probability of the state i at the frame time t by calculating the formula (12) (step S121, FIG. 3). The posterior probability is a value obtained by normalizing the likelihood of each state i (equation (9)) with the sum of the in-frame state likelihood.

The parameter that contributes most to the recognition performance in the speech recognition model is an average vector in the Gaussian distribution. Therefore, in the following description, adaptive learning for the average vector will be described. When focusing on the average vector of the speech recognition model, the auxiliary function Q can be rewritten in the concrete system shown in equation (13).

Here, ζ _{e, k, t} is a posterior probability assigned to the Gaussian distribution k of the speech recognition model e corresponding to the sound source A at the frame time t. The a posteriori probability value ζ _{e, k, t} for each Gaussian distribution is calculated for each Gaussian distribution k in the a posteriori probability calculation unit 121 in the same manner as calculating the a posteriori probability of each state i.

The auxiliary function Q of Expression (13) can be expressed by Expression (14).

Here, 'represents transposition of the matrix. ζ _{e, k} is a sufficient statistic that can be expressed by equation (15) and _{me, k} is expressed by equation (16).

Further, the auxiliary function Q in Expression (14) can be expressed by Expression (17).

Here, μ represents a plurality of speech recognition models handled as one vector as shown in Expression (18).

である。このように複数音声認識モデルの補助関数Ｑは、全音声認識モデルの平均ベクト
ルμの２次形式（式（１７）の右辺第１項）で表現することができるので、安定した解が
得られる。そして、この実施例の適応学習は、初期値音声認識モデルの平均ベクトルμ^０
と推定すべきμに対して式（２１）に示す線形変換を仮定する。 Furthermore,

It is. As described above, the auxiliary function Q of the plurality of speech recognition models can be expressed in the quadratic form of the average vector μ of all speech recognition models (the first term on the right side of the equation (17)), so that a stable solution can be obtained. . The adaptive learning of this embodiment is performed by using the average vector μ ⁰ of the initial value speech recognition model.
Assuming that μ is to be estimated, the linear transformation shown in the equation (21) is assumed.

ここでＢ＝（Ａ，ｂ），ξ＝（（μ^０）´，１）´である。行列Ａは、非対角成分において複数音声認識モデル間のパラメータの相関関係を考慮したものである。

Here, B = (A, b) and ξ = ((μ ⁰ ) ′, 1) ′. The matrix A considers the correlation of parameters between a plurality of speech recognition models in the non-diagonal component.

The relation parameter generation unit 122 substitutes the expression (21) into the expression (17), and a
By calculating rgmax, the parameters A and b are estimated from the adaptive data by the maximum likelihood estimation method (step S122). Parameters A and b correspond to φ shown in equations (4) and (17).

However, since A and b are huge matrices (several hundred thousand dimensions or more), if they are estimated using only adaptive data, the amount of data is insufficient and an overlearning problem occurs. In order to solve this overlearning, the matrix A is blocked and the non-diagonal elements are approximated to zero. Also, by converting b into blocks, the conversion equation (21) can be rewritten as the equation (22).

That is, each average vector μ _{e, k} is converted by A _{e, k} , b _{e, k} . Further, the parameters to be estimated can be further reduced by sharing A and b with a plurality of average vectors. This is achieved by performing clustering on the average vector set in advance and obtaining a cluster including a plurality of average vectors in the cluster according to the data amount. Thereby, A and b can be efficiently estimated with a small number of parameters.

For the clustering of the average vector set for reducing the parameters A and b, a Gaussian distribution shared tree often used in the maximum likelihood linear regression method or the like, which is a typical method for acoustic model adaptation, may be used. The Gaussian distribution shared tree is a technique for expressing a set of Gaussian distributions using a tree structure in which a single Gaussian distribution is a leaf and the set is a node. At this time, an inter-distribution distance such as an Euclidean distance is used to determine which Gaussian distribution is one set.
For example, in the case of a binary tree, two Gaussian distributions having a close distance between distributions are expressed as one node. There are two types of Gaussian distribution tree construction for multiple acoustic models:

(1) By constructing a shared tree using the inter-distribution distance independently for each environment-dependent acoustic model before synthesis, and preparing a common parent node whose root node is a small node,
Synthesize a shared tree. In this case, since the regression matrix is shared within the same speaker, a shared structure using speaker property information is constructed.

(2) A shared tree is constructed by clustering an acoustic model obtained by combining a plurality of models using the distance between distributions. In this case, the regression matrix is shared for Gaussian distributions with close distances between the speakers. That is, it is assumed that speaker property information is not directly taken into account and a phonologically close Gaussian distribution is shared.

In the simulation described later, an adaptation experiment using (2) described above is first performed by combining the methods using two types of shared trees, and the above model is used as the initial model (1).
The adaptation experiment using was conducted.

The update model generation unit 123 calculates the equation (21) using the parameters A and b from the related parameter generation unit 122 and the initial value speech recognition model μ ⁰ recorded in the initial value speech recognition model recording unit 10 as inputs. The voice recognition model is updated (step S123).

  As described above, the speech recognition model creating apparatus 100 according to the first embodiment treats an initial value speech recognition model including a plurality of speech recognition models as one vector, and converts the initial value speech recognition model into a plurality of speech recognition models. It is updated using the speech recognition result recognized by speech recognition based on the state probability transition composed of the combinations of the above. Therefore, since a plurality of speech recognition models can be learned together, a sufficient adaptive learning effect can be obtained even with a small amount of speech data.

[Voice recognition device]
The speech recognition model creation device 100 described in the first embodiment can be used for a speech recognition device. An example of the functional configuration of a speech recognition apparatus 200 using the speech recognition model creation apparatus 100 is shown in FIG. The operation flow is shown in FIG. The speech recognition device 200 includes a speech recognition model creation device 100, a speech recognition network database 22, an A / D conversion unit 91, a feature amount extraction unit 92, a score calculation unit 931, and a speech recognition network selection unit 201. Prepare. The A / D conversion unit 91, the feature amount extraction unit 92, and the score calculation unit 931 are the same as the speech recognition apparatus 900 described in the related art. Therefore, the voice recognition network database 22 and the voice recognition network selection unit 201 will be described.

The voice recognition network database 22 records state probability transitions composed of combinations of a plurality of voice recognition models. This is a recording of a speech recognition network including the expression (5) and a plurality of speech recognition models shown in FIG. Expression (5) means the combination of voice recognition networks when considering the conversation environment of the same language and two speakers. In an environment where words and grammar itself are different as in multilingual speech recognition, each network is prepared and the speech recognition network database 22 is constructed as shown in Expression (23). Expression (23) is the case of transition between utterances. If it is a transition between words, it can be constructed by equation (24).

  In this way, the construction of various acoustic environment models, such as multiple speakers in the same language and transition models between utterances (words) in a multilingual environment, can be realized by synthesizing networks, union operations, etc., which are weighted finite It can be efficiently performed using an existing algorithm such as a state transducer (WFST, a speech recognition decoder using the state transducer is called a WFST type decoder). In the WFST decoder, the acoustic model handles only the information of the HMM state ID and the parameter value of the mixed Gaussian distribution model included therein. Accordingly, regarding the synthesis of a plurality of acoustic models, the parameter values of the mixed Gaussian distribution model corresponding to the ID of the HMM state of each model may be added to the synthesized acoustic model. Attention should be paid to duplication of ID numbers. Also, the corresponding HMM state ID in the WFST needs to be changed accordingly.

  The speech recognition network selection unit 201 has a state probability transition speech recognition in which the score calculated by the score calculation unit 931 using the acoustic feature amount and the updated speech recognition model updated by the speech recognition model creation device 100 is the largest. A state sequence consisting of a network or a set thereof is selected from the speech recognition network database 22 and output as a speech recognition result (step S201). A set of state sequences of speech recognition results is also input to the likelihood calculating unit 13 of the speech recognition model creating apparatus 100, and becomes a teacher signal for adaptive learning.

  The voice recognition network selection unit 201 may output the type e of the voice recognition model constituting the voice recognition network selected together with the state string set as environment information. For example, if the speech recognition network database 22 records two types of speech recognition networks, Japanese e = 1 and English e = 2, the type e is also output. By doing so, there is an effect that it is possible to know the environmental situation of voice recognition.

〔simulation result〕
A simulation was performed for the purpose of confirming the effectiveness of the speech recognition model creation method of the present invention. As the simulation conditions, two types of sex-dependent acoustic models (male and female) were prepared as a plurality of acoustic environments. The speech recognition conditions were a sampling frequency of 16 kHz, a quantization number of 16 bits, a window type of a Hamming window, a frame length of 25 ms, and a frame shift of 10 ms. The language model was trigram (newspaper article 14 years) and the vocabulary was 700,000.

The accuracy of word correctness was compared with the method of the present invention and the conventional method using a single acoustic model independent of gender and a plurality of acoustic models. The results are shown in Table 1.

  The word correct answer rate by the adaptive learning of the present invention showed the best numerical value of 85.5%, and the effect of improving the recognition performance by 1% over the conventional adaptive learning method using a plurality of models was obtained. Compared with single model adaptation, the word accuracy was improved by 3%. As described above, according to the speech recognition apparatus that also uses the speech recognition model creation method of the present invention, an effect of improving the word correct answer accuracy is obtained.

  The speech recognition model creation device and method based on the technical idea of the present invention, and the speech recognition device and method are not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. It is. The processes described in the above-described apparatus and method are not only executed in time series according to the order described, but may be executed in parallel or individually as required by the processing capability of the apparatus that executes the process. . For example, in the above-described embodiment, an example in which the likelihood calculation unit 13 is included in the configuration of the speech recognition model creation device 100 has been described. However, when configuring the speech recognition device, a likelihood calculation unit included in the speech recognition device. The likelihood calculation unit 13 can be deleted by using the likelihood or the score calculated in (1). Further, although an example in which the initial value speech recognition model is recorded as one vector in the initial value speech recognition model recording unit 10 has been described, a plurality of speech recognition models are individually stored in the initial value speech recognition model recording unit 10. And the related parameter generation unit 122 may treat each speech recognition model as one vector. Further, although the voice recognition apparatus 200 has been described as an example including the A / D conversion unit 91, when the voice data is a digitized voice data file, the A / D conversion unit 91 is not necessary.

  Further, when the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape, etc., and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only) Memory), CD-R (Recordable) / RW (ReWritable), etc. can be used as magneto-optical recording media, MO (Magneto Optical disc) can be used, and flash memory can be used as semiconductor memory.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

The figure explaining a union operation notionally. The figure which shows the function structural example of the speech recognition model production apparatus 100 of this invention, and the speech recognition apparatus 200 using the same. The figure which shows the operation | movement flow of the speech recognition model creation apparatus 100. The figure which shows typically 1 state which comprises a phoneme model. The figure which shows an example of a phoneme model. The figure which shows typically the relationship between a flame | frame and the state i. The figure which shows the function structural example of the speech recognition model creation apparatus 100 and the speech recognition apparatus 200. FIG. The figure which shows the operation | movement flow of the speech recognition apparatus 200. The figure which shows the function structural example of the speech recognition apparatus 900 provided with the conventional some speech recognition model.

Claims (10)

An initial value speech recognition model that is a vector obtained by concatenating vectors representing parameters of a plurality of speech recognition models, wherein the plurality of speech recognition models are speech recognition models respectively corresponding to a plurality of sound sources. An initial value speech recognition model recording unit that records a value speech recognition model;
As inputs the feature vectors of a plurality of sets and the audio signal of the state sequence is the result of voice recognition a voice signal voice recognition network based on union operation state transition probability respectively corresponding to each of the speech recognition model, the frame A likelihood calculator that calculates the likelihood of each state for each;
As input and the likelihood and the feature vectors, and the model update unit to generate an updated speech recognition model update the initial value speech recognition models,
An updated speech recognition model recording unit for recording the updated speech recognition model;
A speech recognition model creation device comprising: In the speech recognition model creation device according to claim 1,
The model update unit
A posterior probability calculation unit for calculating a posterior probability value for each Gaussian distribution constituting the state, using the likelihood and the feature vector as inputs,
A relational parameter generating unit that receives the posterior probability value for each Gaussian distribution and the initial value speech recognition model as input, and generates a relational parameter that updates the initial value speech recognition model as one vector;
An update model generation unit that outputs an updated speech recognition model obtained by updating the initial value speech recognition model with the related parameters;
A speech recognition model creation device comprising: In the initial value speech recognition model recording unit, an initial value speech recognition model that is one vector obtained by concatenating vectors representing the parameters of a plurality of speech recognition models corresponding to each sound source is stored.
A set of state sequences as a result of speech recognition of speech signals based on state probability transitions obtained by union-calculating a plurality of speech recognition networks respectively corresponding to the plurality of speech recognition models corresponding to the sound sources ; A likelihood calculation process for calculating the likelihood of each state for each frame , using the feature vector of the speech signal as an input,
Model updating unit, and a model update step of generating an updated speech recognition model on Symbol initial value speech recognition model was updated as input and the likelihood and the feature vector,
An updated speech recognition model recording process in which the updated speech recognition model recording unit records the updated speech recognition model;
Speech recognition model creation method including. In the speech recognition model creation method according to claim 3,
The model update process
A posteriori probability calculation unit calculates a posteriori probability value for each Gaussian distribution constituting the state by using the likelihood as an input; and
A relational parameter generating unit receives a posterior probability value for each Gaussian distribution, the initial value speech recognition model, and a feature vector, and generates a relational parameter for updating the initial value speech recognition model as one vector. Generation step;
An update model generation step, wherein an update model generation unit outputs an updated speech recognition model in which the initial value speech recognition model is updated with the related parameters;
A speech recognition model creation method characterized by comprising: The speech recognition model creation device according to claim 1 or 2,
A speech recognition network database that records state probability transitions consisting of combinations of multiple speech recognition models;
A feature amount extraction unit that extracts a feature amount vector for each frame of a discrete audio signal;
A score calculator that calculates the score using the updated speech recognition model obtained by updating the initial value speech recognition model with the speech recognition result, using the feature vector and the initial value speech recognition model as inputs;
A voice recognition network selection unit that selects the voice recognition network of the state probability transition with the highest score from the voice recognition network database and outputs the result as the voice recognition result;
A speech recognition apparatus comprising: The speech recognition apparatus according to claim 5.
The voice recognition apparatus, wherein the voice recognition network selection unit outputs environment information from the selected voice recognition network. A speech recognition model creation method according to claim 3 or 4,
A feature quantity extraction unit that extracts a feature quantity vector for each frame of a discrete audio signal;
A score calculation process in which a score calculation unit calculates a score corresponding to the feature vector by receiving the feature vector and the updated speech recognition model;
Selecting a speech recognition network of the state probability transition with the highest score from the speech recognition network database and outputting it as a set of state sequences; and
A speech recognition method comprising:   A method program for causing a computer to function the voice recognition model creation method according to claim 3 or 4.   A method program for causing a computer to function the speech recognition method according to claim 7.   A computer-readable recording medium on which the method program according to claim 8 or 9 is recorded.

Images