JP3176210B2 - Voice recognition method and voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method, a speech recognition method, and a speech recognition device, and more particularly, to a speech recognition method and a speech recognition device using tree speaker clustering.

[0002]

2. Description of the Related Art Conventionally, in a continuous speech recognition apparatus, speech recognition is performed using a tree-structured speaker clustering algorithm and the algorithm is adapted to a high-speed speaker adaptation. Tree-structured speaker clustering ", IEICE Technical Report, SP9
3-110, December 1993. In this conventional method, the speaker can be adapted from global features to local features by tracing the clustering tree from a higher level to a lower level. Here, the speaker adaptation is performed by searching the tree structure based on the likelihood of the model for the input speech and selecting the model at the node having the maximum likelihood. That is, speaker adaptation with a teacher signal based on the uttered voice is performed. In this method of speaker adaptation, only the branches of the clustering tree are selected without modifying the parameters.
The advantage is that speaker adaptation can be performed with a small number of samples. In this method, for example, a hidden Markov network that effectively represents a phoneme environment can be used as a phoneme model set.

[0003]

However, the above-described conventional method still has a problem that a sample of adaptation voice data according to the utterance list is still required, and the voice recognition rate is relatively low. An object of the present invention is to solve the above problems, and does not require a sample of audio data for adaptation.
It is an object of the present invention to provide a speech recognition method and a speech recognition device capable of improving the speech recognition rate as compared with the conventional method.

[0004]

According to the first aspect of the present invention, there is provided a voice recognition method comprising a plurality of speaker models stored in a speaker model storage device in advance, and a speech comprising an input character string. In a voice recognition method for recognizing a voice sentence,
The plurality of speaker models stored in advance in the speaker model storage device are such that the top speaker model is an unspecified speaker model and the bottom speaker model is a small number or a specific speaker model. , The clusters are classified hierarchically and expressed in a tree structure, and based on the input uttered speech sentence, speech recognition is performed using an unspecified speaker model including the plurality of speaker models, and the speech recognition is performed. By tracing the tree structure from the uppermost layer to the lowermost layer based on the result and the input uttered speech sentence, the uppermost layer, an intermediate layer located between the uppermost layer and the lowermost layer, and a lowermost layer Selecting at least one more optimal speaker model from among the plurality of speaker models, and re-recognizing the uttered speech sentence based on the selected speaker model, and outputting the speech recognition result. It is characterized by.

According to a second aspect of the present invention, in the speech recognition apparatus, the top speaker model is an unspecified speaker model, and the bottom speaker model is a small number or a specific speaker model. On the basis of a storage device that stores a plurality of speaker models in which the clusters are hierarchically classified and expressed in a tree structure, and a uttered speech sentence composed of an input character string,
Voice recognition means for recognizing voice using an unspecified speaker model consisting of a plurality of speaker models stored in the storage device,
By tracing the tree structure from the uppermost layer to the lowermost layer based on the voice recognition result by the voice recognition unit and the input uttered voice sentence, the uppermost layer and the position between the uppermost layer and the lowermost layer And a selecting means for selecting at least one more optimal speaker model among a plurality of speaker models in the lowermost layer, wherein the speech recognition means comprises:
The utterance speech sentence is again subjected to speech recognition based on the speaker model selected by the selection means, and the speech recognition result is output.

[0006]

[0007]

[0008]

In the speech recognition apparatus constructed as described above, the speech recognition means comprises a plurality of speaker models stored in the storage device based on an uttered speech sentence consisting of an input character string. Speech recognition using an unspecified speaker model. Next, the selecting means traces the tree structure from the uppermost layer to the lowermost layer based on the speech recognition result by the speech recognizing means and the input uttered speech sentence, thereby forming the uppermost layer, the uppermost layer At least one speaker model that is more optimal is selected from among a plurality of speaker models in the middle layer and the lowest layer located between the and the lowest layers. Further, the voice recognition unit performs voice recognition of the uttered voice sentence again based on the speaker model selected by the selection unit, and outputs a voice recognition result.

[0009]

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a speech recognition apparatus according to one embodiment of the present invention, and FIG. 2 is a perspective view showing a configuration of a tree-structure speaker clustering used in the speech recognition apparatus of FIG.

The speech recognition apparatus of the present embodiment performs speech recognition using the conventional tree-structured speaker clustering shown in FIG. 2. In particular, as shown in FIG. 1, a hidden Markov network memory (hereinafter referred to as HM) The phoneme matching unit 4 and a phoneme context-dependent LR parser (hereinafter, referred to as LR parser) based on the unspecified speaker phoneme model stored in the network memory 11.
5, the speech recognition process is executed by a known method, and then the speech recognition result data output from the LR parser 5 is input to the speaker model selection unit 30 as a teacher signal, and is input from the buffer memory 3. HM based on the characteristic parameters of the uttered voice and the voice recognition result data
At least one of the plurality of speaker models in the tree-structured speaker model of the hidden Markov model (hereinafter, referred to as HMM) stored in advance in the network memory 11 is selected, and a selection signal is selected. Output to the HMM memory 11,
In response to this, the phoneme matching unit 4 performs phoneme matching using the HMM of the speaker model corresponding to the selection signal in the HM network memory 11 and performs speech recognition processing.

First, the principle of tree structure speaker clustering will be described with reference to FIG. FIG. 2 shows an example of a speaker model having a hierarchical tree structure.
It has two layered levels. At level 0 of the uppermost layer, there is only one speaker model of the cluster m ⁰ ₀ (0), and this speaker model is a so-called unspecified speaker model. Further, at level 2 of the intermediate layer, the speaker model belongs to m ⁰ ₀ (0) at level 1 and is m ⁰ ₀ (1), which is speaker cluster 1, and speaker cluster 2. m ⁰ ₁ (1) speaker model exists. Further, at the lowest level 3, two speaker clusters m ⁰ ₀ (2) and m belonging to the level 1 speaker cluster 1
⁰ ₁ A speaker model of (2) exists and two speaker clusters m belonging to speaker cluster 2 of level 1
There are speaker models of ⁰ ₀ (2) and m ⁰ ₁ (2). For example, FIG. 2 shows a speaker cluster selected using hatching.
The data of the speaker model of the speaker cluster of ⁰ ₁ (2) is included in the data of the speaker model of the speaker cluster of m ⁰ ₀ (1), which is the speaker cluster 1 of level 1, and the level 0 Speaker model data includes all speaker model data. In other words, as the level number increases, the speaker model is divided into speaker models classified into speaker clusters of more detailed classification. Therefore, the model belonging to the upper layer of the tree structure includes an unspecified number of speaker characteristics, and the model belonging to the lower layer has characteristics of a small number or specific speakers. Although the number of layers is three in the example of FIG. 2, the present invention is not limited to this, and a plurality of layers may be used.

As described above, in the hierarchical speaker clustering, a tree structure of the speaker model is created by sequentially and hierarchically dividing the speaker characteristics. By searching this tree structure based on the likelihood of the model for the input speech, adaptation by speaker selection can be performed. That is, this tree structure is traced from the upper layer to the lower layer, and at least one
By selecting one speaker model, speaker adaptation is possible. If the feature of the input speech has a feature similar to one of the standard speakers constituting the tree structure, it is expected that a lower model is selected. If the characteristics are not similar to those of the standard speaker, it is expected that an upper model is selected. When an upper model is selected, it is considered that an interpolative effect from the characteristics of a plurality of speakers can be obtained.

In order to create a tree structure, a specific speaker phoneme model set for each speaker is first created from data of a plurality of speakers. The plurality of model sets are clustered by a clustering algorithm. The generated individual clusters are further clustered to create subclusters. This is repeated until one cluster becomes one speaker to create a tree structure. After the tree structure is created, a statistical phoneme model set is created from the speech data of the speakers belonging to each cluster. Since a statistical model is used, the likelihood that the model outputs can be used as a criterion for selecting an optimal model. By selecting the model at the node with the highest likelihood,
It is possible to prevent a decrease in robustness. In addition, in the speaker adaptation based on the conventional speaker model selection, it is necessary to increase the number of learning speakers in order to improve the performance. In the method of the present embodiment, the speaker cluster is represented by a tree structure. Thus, the effect of preventing an increase in the amount of calculation required for adaptation when the number of speakers increases can be obtained.

In this embodiment, an HM network is used as a statistical phoneme model set for speech recognition. The HM network is a phoneme environment dependent model expressed efficiently. One HM network includes many phoneme environment dependent models. The HM network is composed of a combination of states including a Gaussian distribution,
State is shared between individual phoneme environment dependent models. Therefore, even when the number of data for parameter estimation is insufficient, a robust model can be created. This HM network is automatically created using a successive state splitting method (SSS). The above S
The SS determines the topology of the HM network, determines the allophone class, and estimates the parameters of the Gaussian distribution in each state at the same time. In the present embodiment, the parameters of the HM network include an output probability and a transition probability expressed by a Gaussian distribution. Therefore, at the time of recognition, it can be handled in the same way as a general HMM.

Further, an algorithm of clustering performed at each node of the tree structure will be described. Here, a method based on the clustering algorithm used in the split (SPLIT) method is used. In this method, a general conventional L that creates a power-of-two cluster
Unlike the BG algorithm, the cluster with the largest distortion is sequentially divided. Therefore, any number of clusters can be created. Before performing clustering, a distance table between elements is created in advance. This has the advantage that the initial value of the cluster center need not be given heuristically (accidentally or heuristically). After all, it is only necessary to give in advance a threshold value for the distance or the number of clusters, and if this value is given, a completely automatic result can be obtained.

A method for creating a tree structure of speaker clusters using the above-described clustering method will be described. In the tree structure creation algorithm proposed here, clusters are automatically created only by giving the number K of clusters at each node. The algorithm is shown below. <Step 1> A plurality of N specific speaker HM networks are created from voice data of a plurality of N speakers. <Step 2> Using the clustering algorithm, clustering of a plurality of N HM networks for fixed speakers is performed to create a plurality of K clusters. Thereafter, the HM network is re-learned using the data of the speakers belonging to each cluster, and a plurality of K HM networks represented by the following equation 1 are created.

[0018]

(1) M ⁰ (j) = {m ⁰ ₀ (j),. . . , M ⁰ _K−1 (j)}, j = 1, 2,. . . , J

Here, j is a hierarchy number indicating the depth of the hierarchy of the tree structure, and here, J is the number of hierarchies.

<Step 3> When the number of speakers s satisfying s∈S ^l (j) becomes equal to or smaller than K, the clustering of the cluster 1 is terminated. Here, S ^l (j) represents the l-th cluster in level hierarchy j. <Step 4> The speakers belonging to the l-th cluster S ^l (j) are clustered for all the l-th clusters except for the cluster completed in step 3, and K sub-clusters are created. Thereafter, the HM network is re-learned based on the data of the speakers belonging to the subcluster, and K HM networks represented by the following Expression 2 are created.

[0021]

## EQU2 ## M ^l (j + 1) = {m ^l ₀ (j + 1),. . . ,
m ^l _K-1 (j + 1)}, j = 1, 2,. . . , J

<Step 5> j is incremented by one. <Step 6> Then, the process returns to step 3.

Further, the principle of unspecified speaker speech recognition by tree-structured speaker clustering according to the present invention will be described. In this embodiment, unsupervised speaker adaptation is performed using evaluation data of only one utterance. The algorithm of the speaker-independent speech recognition method using the tree-structured speaker clustering includes the following steps. <Step 1> The phoneme matching unit 4 and the LR parser 5 recognize the input speech using an unspecified speaker phoneme model. Less than,
The voice recognition in this step is called a first voice recognition process. <Step 2> The phoneme sequence of the recognition result is fed back from the LR parser 5 to the speaker model selection unit 30, and the speaker model selection unit 30 receives the input speech used in step 1 and the speaker Make a selection. <Step 3> Then, the phoneme matching unit 4 and the LR parser 5
Uses the selected phoneme model to perform speech recognition on the input speech again and outputs the result data. Hereinafter, the voice recognition in this step is referred to as a second voice recognition process.

As described above, the final speech recognition result is determined by the two speech recognition processes of the first and second speech recognition processes. In order to improve the speech recognition rate by the speech recognition method of the present embodiment, it is necessary to improve the recognition rate of erroneously recognized data. For this reason, there is an essential problem that, even if an incorrect recognition result is fed back, learning must be performed in the correct direction. However, the speech recognition result data has been corrected to some extent by knowledge of grammar and the like, and when evaluated by a phrase, it is only incorrect and not all phoneme sequences are incorrect. When actually examining the data of the speech recognition error, many of them are incorrect only in the particle part. From this, it is considered that speaker adaptation is sufficiently possible even with feedback of a misrecognition result.

Next, an SSS-LR (left-to-right rightmost type) using the above-described speech recognition method of the present embodiment.
An unspecified speaker continuous speech recognition device will be described. This apparatus uses a phoneme environment-dependent efficient HMM expression format called an HM network stored in a memory 11.
In the SSS, the likelihood of a stochastic model expressing a temporal transition of a speech parameter by a stochastic transition between stochastic stationary signal sources (states) assigned to a feature space of a phoneme is calculated. The refinement of the model is performed sequentially by repeating the operation of dividing each state in the context direction or the time direction based on the criterion of maximization.

In FIG. 1, a speaker's uttered voice is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the phoneme matching unit 4 and the speaker model selecting unit 30 via the buffer memory 3.

HM network memory 1 connected to phoneme matching unit 4
The HM network in 1 is represented as a plurality of networks having each state as a node, and each state has the following information. (A) State number (b) Acceptable context class (c) List of preceding and succeeding states (d) Parameters of output probability density distribution (e) Self transition probability and transition probability to succeeding state

The HM network used in this embodiment is:
Since it is necessary to specify which speaker each distribution originates from, a predetermined speaker mixed HM network is converted and created. Here, the output probability density function is a Gaussian mixture distribution having a 34-dimensional diagonal covariance matrix, and each distribution is learned using a specific speaker sample.

In the first speech recognition process, the phoneme matching unit 4 executes phoneme matching processing in response to a phoneme matching request from the phoneme context-dependent LR parser 5. Then, the likelihood for the data in the phoneme matching section is calculated using the level 0 unspecified speaker model of the uppermost layer shown in FIG. 2, and the value of the likelihood is used as the phoneme matching score in the LR parser 5.
Is returned to Since the model used at this time is equivalent to the HMM, the likelihood calculation uses the forward path algorithm used in the normal HMM as it is.

On the other hand, the context-free grammar database memory 2
As is well known, a predetermined context-free grammar (CFG) in 0 is automatically converted to create an LR table and stored in the LR table memory 13. The LR parser 5 refers to the LR table 13 and processes the input phoneme prediction data from left to right without backtracking. If there is syntactic ambiguity, the stack is split and the analysis of all candidates is processed in parallel. The LR parser 5 predicts the next phoneme from the LR table in the LR table memory 13 and outputs phoneme prediction data to the phoneme matching unit 4. In response to this, the phoneme matching unit 4 sets the H
The collation is performed by referring to the information in the M network memory 11, the likelihood is returned to the LR parser 5 as a speech recognition score, and the continuous speech is recognized by successively connecting the phonemes. Is fed back to the speaker model selection unit 30 and output. When a plurality of phonemes are predicted in the continuous speech recognition, the existence of all of them is checked, and a pruning is performed by using a beam search method to leave a partial tree having a high likelihood of partial speech recognition. To achieve high-speed processing.

Next, in response to this, the speaker model selecting section 30 converts the characteristic parameter data inputted from the buffer memory 3 and the speech recognition result data in the first speech recognition process fed back from the LR parser 5. From the group of speaker models having the tree-structured speaker clustering structure shown in FIG.
A speaker model of a lower speaker cluster having a likelihood equal to or greater than a predetermined threshold is selected, and more preferably, a speaker model of a lowest speaker cluster having a maximum likelihood is selected. Then, a selection signal indicating a speaker cluster of the selected speaker model is output to the HM network memory 11, and a speaker model (hereinafter, referred to as a designated speaker model) used by the phoneme matching unit 4 is designated.

Then, in the second speech recognition process, the phoneme matching unit 4 executes a phoneme matching process in response to a phoneme matching request from the phoneme context-dependent LR parser 5. At this time, the LR parser 5 passes phoneme context information including a phoneme matching section, a phoneme to be matched, and phonemes before and after the phoneme. Based on the received phoneme context information, the phoneme matching unit 4 calculates the likelihood for the data in the phoneme matching section using the specified speaker model,
This likelihood value is returned to the LR parser 5 as a phoneme matching score. In response, the LR parser 5 refers to the LR table 13 and processes the input phoneme prediction data from left to right without regression, as in the first speech recognition process. If there is syntactic ambiguity, the stack is split and the analysis of all candidates is processed in parallel. The LR parser 5 predicts the next phoneme from the LR table in the LR table memory 13 and outputs phoneme prediction data to the phoneme matching unit 4. In response, the phoneme matching unit 4 performs matching by referring to information in the HM network memory 11 regarding the specified speaker model corresponding to the phoneme, and returns the likelihood to the LR parser 5 as a speech recognition score. By successively connecting phonemes, continuous speech is recognized. Here, similarly to the first speech recognition process, when a plurality of phonemes are predicted, the presence of all of them is checked, and a partial tree having a high likelihood of partial speech recognition is determined by a beam search method. High-speed processing is realized by performing pruning to keep. After processing to the end of the input speaker's voice, the one with the highest overall likelihood or a plurality of predetermined higher-order ones is output to an external device as recognition result data of the device.

The following Table 1 shows the results of a simulation performed by the inventor using the speech recognition apparatus of this embodiment.

[0034]

[Table 1]

As is clear from Table 1, although there is a variation depending on the speakers, the method of the embodiment using the tree-structured speaker clustering is the speaker mixing method (for example,
Kosaka et al., "Unspecified Speaker Speech Recognition and Speaker Adaptation Using Mixed SSS," IEICE Technical Report, SP92-
52, September 1992. ), The speech recognition rate is higher than that of unspecified speaker speech recognition. Here, a phrase is used as an evaluation target, but the average time length of the phrase is about 0.9 seconds. Unless an unsupervised speaker adaptation effect is obtained with input speech of this length, unspecified speech is used. It is not expected that the recognition rate will be improved in the user mode. According to experiments, the speech recognition rate has been improved, and it is considered that the effect of speaker adaptation is obtained even if information containing an error in a phrase of 1 second or less is fed back.

As described above, according to this embodiment,
The data of the speech recognition result obtained by the first speech recognition process is fed back to the speaker model selecting unit 30, and a tree structure speech is generated based on the fed back speech recognition result data and the feature parameter input from the buffer memory 3. Selecting a more optimal at least one speaker model from the plurality of speaker models in the HM network having speaker clustering, and executing a second speech recognition process based on the selected speaker model; The result of recognition is output as final recognition result data of the device. As a result, a more optimal speaker model is selected and speech recognition is performed, so that the speech recognition rate is significantly increased. Then, the utterance according to the utterance speaker list becomes unnecessary. That is, it is possible to perform speech recognition at a higher speech recognition rate than in the conventional example without depending on the speaking speaker. In other words, it is possible to provide a speech recognition device that does not require a sample of adaptation speech data and can improve the speech recognition rate as compared with the method of the related art.

A voice recognition method according to the present invention uses at least a plurality of speaker models stored in a speaker model storage device in advance to recognize a voice sentence composed of an input character string. In, based on the input uttered speech sentence, speech recognition using an unspecified speaker model consisting of the plurality of speaker models, based on the speech recognition result and the input uttered speech sentence, It is characterized in that at least one speaker model that is more optimal from a plurality of speaker models is selected, speech recognition is performed based on the selected speaker model, and the speech recognition result is output. The plurality of speaker models stored in advance in the speaker model storage device, which is the HM network memory 11, preferably have at least clusters classified in a hierarchical manner. Further, more preferably, the plurality of speaker models stored in the speaker model storage device in advance may have a cluster represented by a tree structure.

In this embodiment, the speech recognition processing may be executed using the following unsupervised speaker adaptation method. That is, for the input of the adaptation voice,
Speech recognition is once performed by a recognition system, and a phoneme sequence output as a result is fed back and used as a teacher signal at the time of speaker adaptation, whereby apparent unsupervised speaker adaptation can be realized. In this case, in speaker adaptation by tree-structured speaker clustering, only the branch of the tree structure is selected, and parameters such as an average value and a variance are not changed, so that unsupervised learning can be realized with a small amount of data. .

[0039]

As described above in detail, according to the present invention, a plurality of speaker models stored in the speaker model storage device are used in advance to recognize an uttered speech sentence composed of an input character string. In the speech recognition method and apparatus, among the plurality of speaker models stored in the speaker model storage device in advance, the top speaker model is an unspecified speaker model, and the bottom speaker model is a small number or a specified speaker model. As in the speaker model, the clusters are hierarchically classified and represented by a tree structure,
Based on the input uttered speech sentence, speech recognition is performed using an unspecified speaker model including the plurality of speaker models, and based on the speech recognition result and the input uttered speech sentence, the tree structure is used. From the uppermost layer to the lowermost layer, a more optimal at least one story among a plurality of speaker models in the uppermost layer, the intermediate layer located between the uppermost layer and the lowermost layer, and the lowermost layer A speaker model is selected, the uttered speech sentence is again subjected to speech recognition based on the selected speaker model, and the speech recognition result is output. Therefore, by tracing the tree structure of the plurality of speaker models from the uppermost layer to the lowermost layer, the plurality of speaker models in the uppermost layer, the intermediate layer located between the uppermost layer and the lowermost layer, and the lowermost layer Of these, a more optimal speaker model is selected and speech recognition is performed, so that the speech recognition rate is significantly increased. Then, the utterance according to the utterance speaker list becomes unnecessary. That is, there is a specific effect that the voice recognition can be performed at a higher voice recognition rate than the conventional example, without depending on the speaker who utters.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention.

FIG. 2 is a perspective view showing a configuration of a tree-structured speaker clustering used in the speech recognition apparatus of FIG. 1;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3 ... Buffer memory, 4 ... Phoneme collation part, 5 ... LR parser, 11 ... Hidden Markov network memory, 13 ... LR table memory, 20 ... Context-free grammar database memory, 30 ... Talk Part model selection section.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-4-324499 (JP, A) JP-A-61-73199 (JP, A) JP-A-2-220099 (JP, A) JP-A 62-73 245294 (JP, A) JP-A-2-226200 (JP, A) JP-A-59-146099 (JP, A) JP-A-4-20999 (JP, A) (58) Fields investigated (Int. ⁷ , DB name) G10L 3/00 521 G10L 3/00 531

Claims (2)

(57) [Claims] 1. A speech recognition method for recognizing an uttered speech sentence composed of an input character string by using a plurality of speaker models stored in a speaker model storage device in advance. The plurality of speaker models stored in advance are clustered such that the top speaker model is an unspecified speaker model and the bottom speaker model is a small or specific speaker model. Based on the input uttered speech sentence, the speech recognition is performed using an unspecified speaker model including the plurality of speaker models, and the speech recognition result and the input utterance By tracing the tree structure from the uppermost layer to the lowermost layer based on the voice sentence, a plurality of speaker models in the uppermost layer, the intermediate layer located between the uppermost layer and the lowermost layer, and the lowermost layer Out of the more optimal Both selects one speaker model, based on the selected speaker model again recognizing speech the utterance sentence, the speech recognition method and outputs the voice recognition results. 2. The cluster is hierarchically classified and classified into a tree structure, such that the top speaker model is an unspecified speaker model and the bottom speaker model is a minority or specific speaker model. A storage device for storing a plurality of speaker models represented by: and an unspecified speaker model comprising a plurality of speaker models stored in the storage device based on an uttered speech sentence composed of an input character string. Speech recognition means using speech recognition, based on the speech recognition result by the speech recognition means and the input uttered speech sentence, by tracing the tree structure from the top layer to the bottom layer, , An intermediate layer located between the uppermost layer and the lowermost layer, and selecting means for selecting at least one more optimal speaker model among a plurality of speaker models in the lowermost layer, , The above selection means A speech recognition device for re-recognizing the uttered speech sentence based on the speaker model selected in step (a), and outputting the speech recognition result.