JPH0962290A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To find the maximum generation probability for reject judging with the reduced processing amount and the reduced memorizing amount by giving HMM linkage limitation. SOLUTION: A phoneme environmental dependent type phoneme HMM to which phoneme environmental linkage limiting information and language linkage limiting information are added is registered in an HMM data memory 9. A likelihood table preparing part 4 prepares a likelihood table. A reject judging reference cumulative score calculating part 5 calculates a reject judging reference cumulative score along a Viterbi path set on the likelihood table according to the limiting information. A recognition task cumulative score calculating part 6 calculates a recognition task cumulative score. A judging part 7 finds a reject judging maximum reference cumulative score and a recognition task maximum cumulative score to perform reject judgement. As a result, the computing amount of the reject judging maximum reference cumulative score in using the phoneme HMM jointly having a state is reduced, and the memorizing amount of the reject judging maximum reference cumulative score is also reduced.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus using a subword unit HMM (Hidden Markov Model).

[0002]

2. Description of the Related Art Conventionally, HM has been used as a method of speech recognition.
Methods based on M are known. The speech recognition method based on this HMM is described in detail in "Seiichi Nakagawa:" Speech recognition by probabilistic model ", published by The Institute of Electronics, Information and Communication Engineers". The HMM is a kind of state transition network, and the probability of becoming an initial state, the transition probability from state to state, and the output probability of symbols in each state are defined.

Here, the HMM in subword units is the HMM in units smaller than words, such as phonemes and syllabics.
It is MM. HMM for these subword units
Can be applied to word voice recognition, continuous voice recognition, and the like. For example, when the phoneme HMM is used for word speech recognition, the word is regarded as a phoneme string, and it is considered that the concatenation of the phoneme HMM is the feature series of each word. Then, when a voice pattern is input, the characteristic sequence generation probability of each word is obtained by the Viterbi algorithm using the phoneme HMM, and the word with the highest probability is used as the recognition result. On the other hand, when the phoneme HMM is used for continuous speech recognition, a finite state automan syntax network composed of a directed graph of words as shown in FIG. 7 is prepared. This syntactic network can be thought of as a network of phoneme sequences by considering each word as a phoneme sequence. Therefore, the phoneme string having the highest generation probability among the phoneme strings permitted by the syntax network is output as the recognition result by the search process using the phoneme HMM.

When performing word speech recognition or continuous speech recognition as described above, the phoneme HM is usually used for each frame.
The symbol output probability of each state in M or the syntax network is calculated in advance and tabulated. And
When calculating the cumulative score for obtaining the maximum generation rate of each word or the like, a method of referring to the likelihood table created in advance as described above is used.

Here, the voice recognition processing operation requires a reject function which does not output the recognition result when a voice (unknown utterance) not included in the recognition target vocabulary is input. H
For example, "Takao Watanabe, Satoshi Tsukada:" Rejection of unknown utterances by likelihood correction using voice recognition "", The Institute of Electronics, Information and Communication Engineers, D-II, Vol.J75-D-II, N
o.12, pp.2002-2009, December 1992 ”. In this paper, we use semi-syllabic HMM,
It has a syllable network as shown in FIG. 8 independent of the recognition target network. Then, the maximum generation probability obtained from the syllable network and the maximum generation probability obtained from the recognition target network are compared, and if the difference is a certain value or more, it is determined that the utterance content is an unknown utterance. I try to reject it.

[0006]

It is advantageous in terms of processing speed and cost that the processing amount and the storage amount necessary for the rejection determination of the unknown utterance in the voice recognition processing operation are as small as possible. Here, the method of rejecting an unknown utterance with reference to the syllable network as described above is certainly effective. However, it is necessary to store two networks, the syllabic network and the recognition target network, which requires a large storage capacity. Further, since the semi-syllabic HMM is used, there is a constraint that the semi-syllabic HMM for the latter half must be connected next to the semi-syllabic HMM for the first half of the two semi-syllabic parts constituting one syllable. As shown in FIG. 8, there is no restriction on the connection between syllables. Therefore, when calculating the maximum generation probability for reject determination from the syllable network, it is necessary to repeatedly perform a full search of the syllable network regardless of the presence / absence of a semi-syllable HMM that shares a state. The problem is that it requires processing and a large amount of storage.

In addition to the rejection judgment method for unknown utterances referring to the syllable network described above, the paper by Watanabe et al. Describes a simpler rejection judgment method with a smaller amount of processing for each state in each frame. A method is introduced in which the maximum value of the local symbol output probability is accumulated over the entire area of the input, and the accumulated value is treated as a substitute for the maximum generation probability obtained by referring to the syllable network. In addition, this simple method does not place any constraint on the transition of each state, let alone the connection constraint between each semi-syllable HMM, and the state transition probability and the connection condition between semi-syllables (after the VC-type semi-syllable It is also reported that the performance is inferior to the reject judgment method that refers to the syllabic network with the state transition constraint and the linguistic connection constraint because it ignores CV type semi-syllables only). Has been done.

That is, as shown in FIG. 9, the HMM is composed of a number of states connected to each other, and there is a restriction that only a certain state transits to a specific state. Therefore, in the method of accumulating the maximum value of the local symbol output probability of each state in each frame as described above, the constraint of the state transition of the HMM is completely ignored, and the state transition This is accompanied by the problem that the performance is inferior to that of the reject determination method considering the constraints.

Therefore, an object of the present invention is to add a connection constraint of the HMM in addition to the state transition constraint in the HMM so that the maximum generation probability for reject determination can be obtained with a small processing amount and a small storage amount. It is to provide a voice recognition device.

[0010]

In order to achieve the above object, the invention according to claim 1 accumulates an acoustic analysis unit for extracting an acoustic parameter from an input speech and a subword unit HMM having state transition constraint information. Based on the HMM data memory, the extracted acoustic parameters and the stored HMM, the likelihood of creating a likelihood table by calculating the local likelihoods of all the states forming all the HMMs. A probability table creating unit and a route on the likelihood table according to the constraint based on the state transition constraint information of the HMM, and the maximum reference cumulative score along this route is calculated by the Viterbi algorithm. Section and the HMM on the likelihood table
And a recognition task cumulative score calculation unit that calculates a maximum cumulative score along the route according to each recognition task, a maximum reference cumulative score calculated by the reject determination reference cumulative score calculation unit, and the recognition task cumulative score calculation unit The difference between the maximum cumulative score calculated in step 2 is calculated, and if the difference is greater than or equal to a predetermined value, the utterance content is judged to be an unknown utterance that is not recognized and rejected. It has a feature.

In the above structure, when the acoustic analysis unit extracts acoustic parameters from the input voice, the likelihood table creation unit uses the extracted acoustic parameters and the HMMs stored in the HMM data memory. A likelihood table is created. Then, the reject determination reference cumulative score calculation unit sets a route on the likelihood table according to the constraint based on the state transition constraint information of the HMM, and calculates the maximum reference cumulative score along the route by the Viterbi algorithm. To be done. On the other hand, the recognition task cumulative score calculation unit calculates the maximum cumulative score along the route according to the HMM and each recognition task on the likelihood table. Then, the reject determination unit calculates the difference between the two maximum cumulative scores, and if the value of the difference is equal to or greater than a predetermined value, the utterance content is determined to be an unknown utterance that is not a recognition target and is rejected.

At this time, since the maximum reference cumulative score for reject determination is obtained along the Viterbi path set on the likelihood table, when a plurality of HMM states are shared in the same frame, Multiple HMMs above
The calculation path of the maximum cumulative score for reject determination according to is converged into one in the shared state. Therefore, the maximum reference cumulative score for reject determination can be obtained with a small amount of calculation and a small amount of storage.

The invention according to claim 2 is the speech recognition apparatus according to claim 1, wherein the state transition constraint information of the HMM stored in the HMM data memory is H.
It is characterized in that it includes connection constraint information according to the recognition target language at the MM boundary.

According to the above configuration, the route set in the likelihood table is given a constraint in accordance with the connection constraint information according to the recognition target language. Therefore, the maximum reference cumulative score calculation amount for reject determination is calculated. Is even less,
The concatenated HMMs are limited according to the recognition target language, and the maximum reference cumulative score for reject determination is accurately calculated.

According to a third aspect of the present invention, in the voice recognition apparatus according to the first aspect, the HMM stored in the HMM data memory is a phoneme environment-dependent phoneme HMM, and the HMM data memory stores the HMM. HMM
The state transition constraint information of is characterized in that it includes connection constraint information due to the phoneme environment at the HMM boundary.

According to the above configuration, the HMM stored in the HMM data memory is the phoneme environment-dependent phoneme HM.
M, and the route set on the likelihood table is given a constraint according to the connection constraint information due to the phoneme environment, so that the calculation amount of the maximum reference cumulative score for reject determination is further reduced, and The phoneme HMMs to be concatenated are specified by the phoneme environment, and the maximum reference cumulative score for reject determination is calculated very accurately.

[0017]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described in detail below with reference to the embodiments shown in the drawings. FIG. 1 is a block diagram of the speech recognition apparatus according to the present embodiment. This speech recognition device uses an HM obtained from a phoneme HMM and acoustic parameters.
A Viterbi route according to each state transition constraint of the HMM is set on the M likelihood table, and rejection determination of an unknown utterance is performed based on the maximum reference cumulative score on this route.

Before describing the speech recognition apparatus of this embodiment, a maximum cumulative score calculation method along the Viterbi path set on the HMM likelihood table will be described below. FIG. 4 shows how the maximum cumulative score is obtained from the HMM likelihood table by the Viterbi algorithm. Here, FIG. 4A is a path diagram of the state transition when updating the cumulative score in the intermediate state j or the final state m in the state transition shown in FIG. In addition, FIG.
FIG. 10 is a route diagram of state transition when updating the cumulative score in the initial state c in FIG. 9. The vertical axis represents each state, and the horizontal axis represents the frame number.

Each state in a frame of the speech spectrum transits from some states in the frame immediately before that. The state transition route is only the route based on the state transition constraint of the HMM. Also, HMM
Assuming that there is no state sharing between them, there are only two routes that transit to the intermediate state and the final state in the HMM as shown in FIG. In FIG. 4, the state j on the i frame in FIG. 4A is transited from the state c and the state j on the immediately preceding (i−1) frame. The maximum cumulative score among the cumulative scores of both routes from the (i-1) frame to the state j on the i frame is the cumulative score of the state j on the i frame. The above cumulative score is the formula
It is calculated by the Viterbi algorithm shown in (1).

[Equation 1]

As described above, for each frame, equation (1)
The cumulative score of each state is calculated by, but at that time, it is not necessary to store the route that has reached the state. In addition, the calculation of the cumulative score in each state is
It is possible if the cumulative score in each state on the immediately preceding frame is held. Therefore, only the cumulative score in each state on the immediately preceding frame needs to be stored, and the required storage amount can be reduced.

As shown in FIG. 9, the HMM goes from an initial state to an intermediate state to a final state. Then, after leaving the final state, the state transition is newly started from the initial state of another HMM or the initial state of the same HMM. Here, in the case of an HMM of a type that does not depend on the preceding phoneme or syllable or the subsequent phoneme or syllable, the number of initial states of the next HMM that can be connected from the final state of the previous HMM is large. FIG. 4B shows an initial state c on the i frame.
FIG. 7 shows a transition from the initial state c of the same HMM and the final states a, f, m, p of another HMM on the (i-1) frame. In that case, in terms of language, it is possible to set restrictions on the connection of specific phonemes and phonemes or syllables and syllables. For example, in Japanese, it is considered that consonants and consonants are often not connected.
It is possible to set a connection constraint (hereinafter referred to as a linguistic connection constraint) depending on the recognition target language such as not providing a transition path from m / to phoneme / h /.

FIG. 5 and FIG. 6 show examples of connection constraints between phonemes in Japanese, which is the above-mentioned linguistic connection constraint, and show that the final state in the left phoneme HMM can transit to the initial state in the right phoneme HMM. I mean. However,
The phoneme notation follows the Hebon-type Roman spelling, where / q / represents consonant and / N / represents sound repellency. Therefore, in the example shown in FIG. 5, the phonemes connectable to the phoneme / h / are / a /, / i /, / u.
Since there are 7 phonemes of /, / e /, / o /, / N /, / q /, the phoneme / h /
There are seven routes that can transit to the initial state of. In this way, by specializing in the recognition language, it becomes possible to more accurately obtain the cumulative score.

The phoneme environment-dependent phoneme HMM is a specification of the preceding phoneme and the subsequent phoneme related to each phoneme. For example, even if the same phoneme / a /, the preceding phoneme is / k /
The phoneme / a / is a phoneme different from the phoneme / a / in which the preceding phoneme is / m /, and each phoneme is represented by a different phoneme HMM. The phoneme environment-dependent phoneme HMM requires more models than the phoneme environment-independent phoneme HMM, but is advantageous in performance. As described above, in the case of the phoneme environment-dependent phoneme HMM, since the connection constraint between the phoneme HMMs is as important as the transition constraint between states, the phoneme environment-dependent phoneme is used in the calculation of the cumulative score for reject determination. It can be said that giving a connection constraint between HMMs (hereinafter referred to as a phoneme environment connection constraint) is particularly effective in terms of accuracy and ease of calculation of the cumulative score.

As the cumulative score finally referred to in the rejection judgment, the cumulative score which exhibits the maximum value in the final frame is used. At that time, the initial state and the intermediate state may be excluded, and the cumulative score having the maximum value may be obtained from the cumulative scores related to the final state.

A specific example of the voice recognition device to which the above-described maximum cumulative score calculation method is applied will be described below with reference to FIG. In FIG. 1, the audio signal input from the microphone 1 is A / D converted by the A / D converter 2 and sent to the acoustic analysis unit 3. Then, the acoustic analysis unit 3 extracts acoustic parameters for each frame based on the digital audio data. The likelihood table creation unit 4 uses the extracted acoustic parameters and all the phoneme HMMs stored in the HMM data memory 9 to generate the symbol output probabilities of all the different states forming the whole phoneme HMM. o is calculated and a likelihood table is created. Then,
The reject determination reference cumulative score calculation unit 5 is shown in FIG. 4 in which the linguistic connection constraint and the phoneme environment connection constraint are added as described above using the created likelihood table and the phoneme HMM. The reference cumulative score for reject determination in each state on such a route is calculated for each frame, and the content of the reference cumulative score for reject determination storage unit 10 is updated. On the other hand, the recognition task cumulative score calculation unit 6
Uses the likelihood table, each phoneme HMM in the HMM data memory 9 and the recognition task dictionary 12 to calculate a recognition task cumulative score which is a cumulative value of the recognition task generation probability for each frame, and stores the recognition task cumulative score. Update the contents of section 11. When the above-mentioned processing for the final frame is completed, the determination unit 7 stores the maximum value of the reference cumulative score for reject determination stored in the reference cumulative score for rejection determination 10 and the stored value in the recognition task cumulative score storage unit 11. The recognition result is output or rejected based on the normalized value of the difference from the maximum value of the accumulated recognition task score. The voice recognition control unit 8 controls the A / D converter 2, the acoustic analysis unit 3, the likelihood table creation unit 4, the reject determination reference cumulative score calculation unit 5, the recognition task cumulative score calculation unit 6, and the determination unit 7. Voice recognition processing.

Here, the phoneme HMMs stored in the HMM data memory 9 are phoneme environment dependent phoneme HMMs, and phoneme environment connection constraint information and linguistics as shown in FIGS. 5 and 6 are provided at the HMM boundaries. Connection constraint information is added.

FIG. 2 shows that, under the control of the voice recognition control section 8, the A / D converter 2, the acoustic analysis section 3, the likelihood table creation section 4, the reject decision reference cumulative score calculation section 5,
6 is a flowchart of a voice recognition processing operation executed by a recognition task cumulative score calculation unit 6 and a determination unit 7. Hereinafter, the voice recognition processing operation according to the present embodiment will be described with reference to FIG.

In step S1, the reject determination reference cumulative score (hereinafter simply referred to as reject cumulative score) gr _j (j: state number) stored in the reject determination reference cumulative score storage unit 10 is initialized. It In step S2, a recognition task cumulative score (hereinafter simply referred to as a recognition cumulative score) gt _K (k: recognition task number) stored in the recognition task cumulative score storage unit 11 is initialized. In step S3, the frame number i and the frame number I are initialized to "0". In step S4, the frame number i is incremented. In step S5, the audio signal of the i-th frame of the audio signal input from the microphone 1 is captured. In step S6, the audio signal of the frame i is A / D converted and digitized by the A / D converter 2. In step S7, the acoustic analysis unit 3 extracts the acoustic parameter of the frame i from the digital audio signal.

In step S8, the likelihood table creating unit 4 uses the extracted acoustic parameters of the frame and the phoneme HMM data of the HMM data memory 9 to extract all phonemes HM stored in the HMM data memory 9.
The symbol output probabilities o _j (i) of all the different states j forming M are calculated. Here, the above phoneme HMM
The data of is generated by the task-independent learning using the phoneme-balanced words of many speakers. In step S9, the likelihood table creation unit 4 performs normalization or logarithmic transformation on the calculated symbol output probability o _j (i) of each state to obtain the likelihood of the state. The corresponding frame of the frequency table is created. Here, the likelihood table has a configuration as shown in FIG. 3, and the likelihood is stored for each state in each frame. Actually, it is not necessary to store this likelihood table in another storage unit, but the relevant frame is held in the internal memory of the likelihood table creation unit 4 or the like, and the reject cumulative score gr in the relevant frame i is stored. _It may be deleted after the calculation of _j (i) and the recognition cumulative score gt _k (i) is completed. In this way, the amount of memory required for the voice recognition processing operation is reduced.

In step S10, the immediately preceding frame stored in the reject determination reference cumulative score storage unit 10 by the reject determination reference cumulative score calculation unit 5
Reject cumulative score gr _j in each state j of (i-1)
(i-1) is read out, and each phoneme HMM stored in the HMM data memory 9 and the likelihood table are used to follow the state transition path based on each phoneme HMM as shown in FIG. The reject cumulative score gr _j (i) in each state j of i is calculated by the equation (1). Then, the reject cumulative score gr _j (i) thus calculated
Accordingly, the reference cumulative score storage unit 10 for reject determination
Is updated.

In other words, the reject determination reference cumulative score calculation unit 5 adds the phoneme H to the likelihood table.
Each state j along the Viterbi path set according to MM
The reject cumulative score gr _j (i) is calculated for each. At that time, the phoneme HMM stored in the HMM data memory 9 is a phoneme environment-dependent phoneme HMM,
Phoneme environment connection restriction information and linguistic connection restriction information are added to the HMM boundary. Therefore, the state transition paths set on the likelihood table based on each phoneme HMM are limited, and the number of calculation processes of the reject cumulative score gr _j (i) by the Viterbi algorithm is reduced. .

In step S11, the recognition task cumulative score storage unit 11 is operated by the recognition task cumulative score calculation unit 6.
The recognition cumulative score gt _k (i-1) of each recognition task k in the immediately preceding frame (i-1) stored in
The recognition cumulative score gt _k (i) is calculated by using each phoneme HMM stored in the MM data memory 9, the likelihood table and the recognition task dictionary 12. Then, the content of the recognition task cumulative score storage 11 is updated with the recognition cumulative score gt _k (i) thus calculated. Here, the recognition task is represented by concatenation of phoneme HMMs. Therefore, the symbol output probability o _k (i) and the state transition probability p _k (i) in the frame i are added to the recognition cumulative score gt _k (i-1) of each recognition task k in the previous frame (i-1). By accumulating the recognition cumulative score gt _{k of the} frame i.
(i) is calculated.

In other words, the recognition task cumulative score calculation unit 6 recognizes the recognition cumulative score gt _{k for} each recognition task k along the Viterbi path set according to the phoneme HMM and the recognition task on the likelihood table. (i) is calculated.

In step S12, it is judged whether or not the frame i is the final frame. As a result, if it is not the final frame, the process returns to the step S4 to shift to the processing of the next frame (i + 1). On the other hand, if it is the final frame, the process proceeds to step S13. In step S13, the frame number i is set to the frame number I. Step S14
Then, the judging unit 7 searches the reject accumulating score gr _j (I) stored in the reject judging reference accumulative score storing unit 10 for the maximum value and sets it as the reject judging maximum reference accumulating score Lr. . Similarly, the maximum value is retrieved from the recognition cumulative score gt _k (I) stored in the recognition task cumulative score storage unit 11 and set as the recognition task maximum cumulative score Lt.

In step S15, the judging section 7
The maximum reference cumulative score L for reject judgment retrieved above
From r and the recognition task maximum cumulative score Lt, the normalized reject determination value L ′ is calculated by the equation (2). L '= (Lr-Lt) / I (2) Here, the maximum reference cumulative score Lr for reject determination is given.
And the recognition task maximum cumulative score Lt is the number of frames I
It increases in proportion to. Therefore, both cumulative scores Lr,
The difference in Lt is normalized by the number of frames I. In step S16, the determining unit 7 determines whether the calculated normalized reject determination value L'is smaller than a threshold value. As a result, if it is smaller than the threshold value, the process proceeds to step S18, and if it is equal to or larger than the threshold value, step S17.
Proceed to. In step S17, the determination unit 7 determines that the utterance content is an unknown utterance, rejects the phoneme sequence having the maximum recognition task reference score Lt, and ends the voice recognition processing operation. In step S18, the determination unit 7 determines that the utterance content is included in the recognition vocabulary, outputs the phoneme sequence having the maximum recognition task reference score Lt as the recognition result, and ends the voice recognition processing operation. To do.

As described above, in the present embodiment,
In the HMM data memory 9, a phoneme environment-dependent phoneme HMM in which phoneme environment connection constraint information and linguistic connection constraint information are added to HMM boundaries is registered. Then, the likelihood table creating unit 4 creates a likelihood table from the acoustic parameters based on the audio signal captured for each frame and the phoneme HMM in the HMM data memory 9. Then, the reject determination reference cumulative score calculation unit 5 causes the phoneme H to which the phoneme environment connection constraint and the linguistic connection constraint are added.
The reject cumulative score gr _j (i) is calculated by the Viterbi algorithm along the route set on the likelihood table according to the MM, and the content of the reject determination reference cumulative score storage unit 10 is updated. On the other hand, the recognition task cumulative score calculation unit 6 adds the phoneme HMM on the likelihood table.
Then, the recognition cumulative score gt _k (i) is calculated for each recognition task k along the route set according to the recognition task, and the contents of the recognition task cumulative score storage unit 11 are updated. When the above process is completed up to the final frame, the judgment unit 7 determines that both cumulative scores gr _j (i), stored in the reject judgment reference cumulative score storage unit 10 and the recognition task cumulative score storage unit 11 at that time. The maximum value of gt _k (i) is searched to obtain the maximum reference cumulative score Lr for reject determination and the maximum cumulative score Lt of the recognition task. Then, the rejection judgment of the recognition result is performed by the normalized reject judgment value L ′ obtained based on both the maximum cumulative scores Lr, Lt.

As described above, the reject determination reference cumulative score calculation unit 5 in the present embodiment, reject cumulative score gr for each state j along the Viterbi path set according to the phoneme HMM on the likelihood table. _j (i) is calculated. Therefore, when a certain state j in the frame is shared by, for example, two phoneme HMMs, two paths that transit from a different state in the immediately preceding frame in the two phoneme HMMs to a shared state j in the frame. One of them (the one with the smaller reject cumulative score) is cut off at the frame by the Viterbi algorithm. As a result, it is not necessary to calculate the reject cumulative score gr _j (i) for the disconnected phoneme HMM in the subsequent frames, and the amount of calculation can be reduced accordingly. On the other hand, in the “likelihood correction method using syllable recognition” described in the paper of Watanabe et al. Described in the section of the conventional art, the generation probability is calculated using a syllable network as shown in FIG. It is calculated. Therefore, even if there is partial state sharing among the syllable HMMs, the cumulative score is obtained independently for each syllable HMM. Therefore, the amount of calculation of the cumulative score is not reduced. That is, according to the present embodiment, when the phoneme HMM to be used includes the state-shared phoneme HMM, the processing amount at the time of reject determination is significantly reduced.

Further, since the phoneme environment connection constraint information and the linguistic connection constraint information are added to the boundary of the phoneme HMM, the amount of calculation processing can be further reduced by limiting the phoneme HMMs to be connected. Furthermore, it is possible to accurately calculate the reject cumulative score gr _j (i) by deleting the impossible route in advance, and to improve the rejection determination accuracy.

The calculation of the reject cumulative score gr _j (i) and the calculation of the recognition cumulative score gt _k (i) are both performed by the HMM.
It is created based on the phoneme HMM data accumulated in the data memory 9. Therefore, unlike the prior art, it is not necessary to store two networks for reject determination and recognition task, and the storage amount can be reduced. Further, in the above embodiment, the reject cumulative score gr _j is calculated from the phoneme HMM stored in the HMM data memory 9.
Since (i) and the cumulative recognition score gt _k (i) are calculated, the above-described reject determination process can be applied regardless of whether the recognition task is a word or continuous speech recognition using a syntactic network.

In the above embodiment, the HM
A phoneme environment-dependent phoneme HMM is registered in the M data memory 9, and phoneme environment connection restriction information and linguistic connection restriction information are added to the HMM boundaries as restriction information. However, the present invention is not limited to this, and only the phoneme environment connection constraint information may be added as the constraint information. In addition, it does not matter even if the non-phoneme environment-dependent phoneme HMM is registered in the HMM data memory and only the linguistic connection constraint information is added to the HMM boundary. However, in that case, the accuracy of calculating the reject cumulative score gr _j (i) becomes low.

[0041]

As is apparent from the above, the speech recognition apparatus of the invention according to claim 1 creates a likelihood table by the likelihood table creating section based on the acoustic parameter and the HMM in units of subwords for reject judgment. The reference cumulative score calculation unit calculates the maximum reference cumulative score along the route set on the likelihood table by the Viterbi algorithm according to the constraint based on the state transition constraint information of the HMM, and the recognition task cumulative score calculation unit calculates the maximum reference cumulative score. The maximum cumulative score along the route according to the HMM and each recognition task on the likelihood table is calculated, and the reject determination unit determines the rejection of the utterance content based on the difference between the maximum cumulative scores. When calculating the maximum reference cumulative score for judgment, multiple HMM states are shared in the same frame. When the operation amount of the maximum reference cumulative score is greatly reduced. As a result, the storage amount of the maximum reference cumulative score can be reduced. In addition, the maximum reference cumulative score for reject judgment and the maximum cumulative score for recognition task are the same for HMM in the same subword unit.
Since it is calculated from the likelihood table based on, it is not necessary to separately provide networks for reject determination and recognition tasks, and the storage capacity can be reduced.

Further, since the state transition constraint information of the HMM in the speech recognition apparatus according to the second aspect includes the connection constraint information depending on the recognition target language at the HMM boundary, the route set on the likelihood table is The above linguistic connection constraint is added. Therefore, according to the present invention, in addition to the effect of the invention according to claim 1, the operation amount of the maximum reference cumulative score for reject determination is further reduced, and the concatenated HMMs are limited according to the recognition target language. Therefore, the accuracy of calculation of the maximum reference cumulative score for reject determination is improved.

The HMM stored in the HMM data memory in the speech recognition apparatus according to the third aspect is a phoneme environment-dependent phoneme HMM, and the state transition constraint information of the HMM is the phoneme environment at the HMM boundary. Since the connection constraint information is included, the phoneme environment connection constraint is given to the route set on the likelihood table. Therefore, according to the present invention, in addition to the effect of the invention according to claim 1, the operation amount of the maximum reference cumulative score for reject determination is further reduced, and the concatenated phoneme HMM is identified by the phoneme environment and rejected. The accuracy of calculation of the maximum reference cumulative score for determination becomes very good.

[Brief description of drawings]

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

FIG. 2 is a flowchart of a voice recognition processing operation performed under the control of a voice recognition control unit in FIG.

FIG. 3 is a diagram showing an example of a likelihood table.

FIG. 4 is a diagram showing an example of a route for obtaining a cumulative score from a likelihood table by a Viterbi algorithm.

FIG. 5 is a diagram showing an example of a connection constraint between phonemes.

FIG. 6 is a diagram showing an example of a connection constraint between phonemes subsequent to FIG. 5;

FIG. 7 is a diagram showing an example of a syntax network.

FIG. 8 is a diagram showing a syllable network used in conventional unknown speech rejection.

FIG. 9 is an explanatory diagram of an HMM.

[Explanation of symbols]

3 ... Acoustic analysis unit, 4 ... Likelihood table creation unit, 5 ... Rejection determination reference cumulative score calculation unit,
6 ... Recognition task cumulative score calculation unit, 7 ... Judgment unit, 8 ...
Speech recognition control unit, 9 ... HMM data memory, 10 ... Reject judgment reference cumulative score storage unit, 1
1 ... Recognition task cumulative score storage unit.

Claims (3)

[Claims] 1. An acoustic analysis unit for extracting an acoustic parameter from an input voice, an HMM data memory in which a sub-word unit HMM having state transition constraint information is accumulated, the extracted acoustic parameter and the accumulated acoustic parameters. H
A likelihood table creation unit that creates a likelihood table by calculating local likelihoods of all the states that compose all HMMs based on the MM, and the likelihood table creation unit according to the constraint based on the state transition constraint information of the HMM. A reference determination score calculation unit for reject determination that sets a route on the frequency table and calculates the maximum reference accumulation score along this route by the Viterbi algorithm, and a route that follows the HMM and each recognition task on the likelihood table. Of the recognition task cumulative score calculation unit that calculates the maximum cumulative score according to, the maximum reference cumulative score calculated by the reference cumulative score calculation unit for reject determination, and the maximum cumulative score calculated by the recognition task cumulative score calculation unit The difference is calculated, and if the value of this difference is greater than or equal to a predetermined value, it is determined that the utterance content is an unknown utterance that is not a recognition target and rejected. Speech recognition apparatus comprising the object determination unit. 2. The voice recognition device according to claim 1, wherein the HMM state transition constraint information stored in the HMM data memory includes connection constraint information in a recognition target language at an HMM boundary. Voice recognition device. 3. The speech recognition apparatus according to claim 1, wherein the HMM stored in the HMM data memory is a phoneme environment-dependent phoneme HMM, and the state of the HMM stored in the HMM data memory. The speech recognition apparatus, wherein the transition constraint information includes connection constraint information depending on a phoneme environment at an HMM boundary.