JP3315565B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a speech recognition apparatus using a HMM (Hidden Markov Model) in subword units.

[0002]

2. Description of the Related Art Conventionally, HM has been used as one method of speech recognition.
Methods based on M are known. The speech recognition method based on the HMM is described in detail in "Seiichi Nakagawa:" Speech Recognition by Stochastic Model ", published by IEICE." The HMM is a kind of a state transition network, and defines a probability of an initial state, a transition probability from a state to a state, and a symbol output probability in each state.

[0003] Here, the HMM in sub-word units refers to HMM in units smaller than words, such as phonemes or semi-syllables.
MM. HMM of these subword units
Can be applied to word speech recognition, continuous speech recognition, and the like. For example, when a phoneme HMM is used for word speech recognition, a word is regarded as a phoneme sequence, and a combination of phoneme HMMs is considered to be a feature sequence of each word. Then, when a speech pattern is input, a feature sequence generation probability of each word is obtained by a Viterbi algorithm using a phoneme HMM, and the word having the highest probability is used as a 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 considered as a phoneme network by considering each word as a phoneme sequence. Therefore, the search processing using the phoneme HMM outputs the phoneme string having the highest generation probability among the phoneme strings permitted by the syntax network as a recognition result.

[0004] When performing the above-described word speech recognition or continuous speech recognition, usually, the phoneme HM is 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 adopted.

[0005] Here, the speech recognition processing operation requires a reject function of not outputting a recognition result when speech (unknown utterance) not included in the vocabulary to be recognized is input. H
Regarding the method of rejecting unknown utterances in the MM speech recognition device, for example, “Takao Watanabe, Satoshi Tsukada:“ Rejection of unknown utterances by likelihood correction using speech recognition ”, IEICE Transactions, D-II, Vol.J75-D-II, N
o.12, pp.2002-2009, December 1992 ". In this paper, we use a semi-syllable HMM,
It has a syllable network as shown in FIG. 8 independently of the recognition target network. Then, the maximum generation probability obtained from the syllable network is compared with the maximum generation probability obtained from the recognition target network, and if the difference is equal to or greater than a certain value, the utterance content is determined to be an unknown utterance. I try to reject.

[0006]

In the speech recognition processing operation, the smaller the processing amount and storage amount required for the rejection determination of the unknown utterance, the more advantageous in terms of processing speed and cost. 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 syllable network and the network to be recognized, which requires a large storage capacity. In addition, since a semisyllable HMM is used, there is a restriction that a half-syllable HMM for the second half must be connected to a half-syllable HMM of the first half related to two half-syllables 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 rejection determination from the syllable network, it is necessary to repeatedly perform a full search of the syllable network regardless of the presence or absence of a semisyllable HMM sharing a state. There is a problem of requiring processing and a large amount of storage.

In the paper by Watanabe et al., In addition to the rejection determination method for unknown utterances referring to the syllable network described above, a simpler rejection determination method with a reduced processing amount is described in each state of 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 this accumulated value is used as a substitute for the maximum generation probability obtained by referring to the syllable network. In this simplified method, there is no restriction on the transition of each state as well as the connection restriction between the semi-syllable HMMs, and the state transition probability and the connection condition between the semi-syllables (after the VC type semi-syllable, Because only CV type syllables can be connected, etc.), it is reported that the performance is inferior to the reject decision method that refers to a syllable network with state transition constraints and linguistic connection constraints. Have been.

That is, as shown in FIG. 9, the above HMM is composed of a connection of several states, and there is a restriction that a transition from a certain state to only a specific state is made. Therefore, in the method of accumulating the maximum value of the local symbol output probability of each state in each frame over the entire area of the input as described above, the constraint of the state transition of the HMM is completely ignored, and the state transition of the state transition is completely ignored. There is a problem that the performance is inferior to that of the reject determination method in consideration of the restriction.

Therefore, an object of the present invention is to provide 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. A voice recognition device is provided.

[0010]

According to a first aspect of the present invention, there is provided an audio analysis unit for extracting an audio parameter from an input voice, and an HMM in subword units having state transition constraint information is stored. Based on the HMM data memory, the extracted acoustic parameters and the stored HMMs, local likelihoods of all states constituting all the HMMs are calculated to generate a likelihood table. Calculating a reference cumulative score for reject determination by setting a route on the likelihood table according to a constraint based on the state transition constraint information of the HMM and calculating a maximum reference cumulative score along the route by a Viterbi algorithm And the HMM on the likelihood table
A recognition task cumulative score calculation unit that calculates a maximum cumulative score along a path 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 Calculating a difference from the maximum cumulative score calculated in the above, 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 not to be recognized and is rejected.
Of the HMM stored in the HMM data memory.
The state transition constraint information depends on the language to be recognized at the HMM boundary.
It is characterized by including connection constraint information between phonemes .

[0011] In the above configuration, when acoustic parameters are extracted from the input speech by the acoustic analysis unit, the likelihood table creating unit creates the likelihood based on the extracted acoustic parameters and the HMM stored in the HMM data memory. A likelihood table is created. A route is set by the reject determination reference cumulative score calculation unit 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 the route is calculated by the Viterbi algorithm. Is done. On the other hand, the recognition task cumulative score calculation unit calculates the maximum cumulative score along the path 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 not to be recognized and rejected.

At this time, since the maximum reference cumulative score for the reject determination is obtained along the Viterbi path set on the likelihood table, when the state of a plurality of HMMs is shared in the same frame, The above multiple HMMs
The calculation path of the maximum cumulative score for reject determination according to the above is converged to one in the above-mentioned 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.

[0013]

Furthermore, since the path set on the likelihood table constraint in accordance with the communication <br/> binding constraint information among phonemes by the recognition target language in HMM boundary is given, for rejection determination calculating the amount of the maximum reference cumulative score with there is less, the maximum reference cumulative score for rejection determination HMM coupled is limited in accordance with the recognition target language is accurately calculated.

According to a second aspect of the present invention, the input voice
Sound analysis unit that extracts sound parameters from the
HMMs in sub-words with information about
HMM data memory and the extracted acoustic parameters
Based on the data and the stored HMM,
Calculate the local likelihood of all the constituent states and calculate the likelihood table.
Likelihood table creation unit for creating a table,
The likelihood table according to the constraint based on the state transition constraint information
Above, and the maximum reference cumulative score along this path
Rejection calculation that calculates a by the Viterbi algorithm
The reference cumulative score calculation unit and the likelihood table
Along the path according to the above HMM and each recognition task
Recognition task cumulative score calculation to calculate the maximum cumulative score
Unit and the reference cumulative score calculation unit for reject determination described above.
The maximum reference cumulative score issued and the cumulative score
A) Calculate the difference between the maximum cumulative scores calculated by the
If the value of the difference is equal to or greater than the predetermined value,
Reject decision to judge as unknown utterance and reject
It comprises a tough, HMM stored in the HMM data memory is a phoneme HMM phoneme environment-dependent, the H
The HMM state transition constraint information stored in the MM data memory is characterized in that it includes connection constraint information based on the phoneme environment at the HMM boundary.

According to the above configuration, the HMM stored in the HMM data memory is a phoneme environment-dependent phoneme HM.
M, and the path set on the likelihood table is given a constraint according to the connection constraint information based on the phoneme environment, so that the calculation amount of the maximum reference cumulative score for reject determination is further reduced, The connected phoneme HMM is specified by the phoneme environment, and the maximum reference cumulative score for reject determination is calculated with very high accuracy.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments. FIG. 1 is a block diagram of the speech recognition apparatus according to the present embodiment. This speech recognition apparatus uses an HM obtained from a phoneme HMM and acoustic parameters.
A Viterbi route is set on the M likelihood table along each state transition constraint of the HMM, and reject determination of an unknown utterance is performed based on the maximum reference cumulative score on this route.

Before describing the speech recognition apparatus according to the present embodiment, a method of calculating the maximum cumulative score along the Viterbi path set on the HMM likelihood table will be described. FIG. 4 shows how the maximum cumulative score is obtained from the HMM likelihood table by the Viterbi algorithm. Here, FIG. 4A is a state transition path diagram when updating the accumulated score in the intermediate state j or the final state m in the state transition shown in FIG. FIG. 4 (b)
FIG. 10 is a path diagram of state transition when updating a cumulative score in an initial state c in FIG. 9. The vertical axis indicates each state, and the horizontal axis indicates the frame number.

Each state in a certain frame of the voice spectrum transitions from some states in the immediately preceding frame. The path of this state transition is only a path based on the state transition constraint of the HMM. Also, HMM
Assuming that there is no state sharing between the HMMs, there are only two paths that transition between the intermediate state and the final state in the HMM as shown in FIG. In FIG. 4, a transition is made from the state c and the state j on the immediately preceding (i-1) frame to the state j on the i-th frame in FIG. Then, among the cumulative scores of the two paths from the (i-1) frame to the state j on the i frame, the largest cumulative score is the cumulative score of the state j on the i frame. The cumulative score is calculated by 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 according to the above, but at this time, it is not necessary to store the path that has been reached to the state. The calculation of the cumulative score in each state is as follows:
It is possible if the accumulated score in each state on the immediately preceding frame is held. Therefore, only the accumulated 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 the initial state to the final state via the intermediate state. Then, after exiting the final state, a state transition is newly started from another HMM initial state or the same HMM initial state. Here, in the case of a type of HMM that does not depend on the preceding phoneme or syllable or the succeeding 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 the initial state c on the i-frame.
FIG. 7 (b) shows a transition from the initial state c of the same HMM and the final states a, f, m, p of the other HMMs on the (i-1) frame. In this case, it is possible to limit the connection between specific phonemes and phonemes or between syllables and syllables in terms of language. For example, in the case of Japanese, consonants and consonants are often not connected,
It is possible to set a connection constraint (hereinafter referred to as a linguistic connection constraint) based on the recognition target language, such as not providing a transition path from m / to a phoneme / h /.

FIG. 5 and FIG. 6 show an example of a connection restriction between phonemes in Japanese, which is the above-mentioned linguistic connection restriction, and show that the transition from the final state in the left phoneme HMM to the initial state in the right phoneme HMM is possible. Means. However,
The phoneme notation follows the Hepburn roman spelling, where / q / indicates a prompting sound and / N / indicates a sound-repelling sound. Therefore, in the example shown in FIG. 5, phonemes connectable to phoneme / h / are / a /, / i /, / u
/, / e /, / o /, / N /, / q /
There will be seven paths that can transition to the initial state. In this way, by specializing in the recognition language, it is possible to more accurately calculate the cumulative score.

The phoneme environment-dependent phoneme HMM specifies a preceding phoneme and a succeeding phoneme relating to each phoneme. For example, even if the same phoneme is / a /, the preceding phoneme is / k /
Is treated as a different phoneme from the phoneme / a / whose preceding phoneme is / m /, and each phoneme is represented by a separate 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 a phoneme environment-dependent phoneme HMM, since the connection constraint between phoneme HMMs is as important as the transition constraint between states, the phoneme by the phoneme environment is used in calculating the cumulative score for reject determination. It can be said that providing a connection constraint between HMMs (hereinafter referred to as a phoneme environment connection constraint) is particularly effective in terms of the accuracy and ease of calculating the cumulative score.

Incidentally, as the cumulative score finally referred to at the time of the reject determination, the cumulative score that exhibits the maximum value in the final frame is used. At this time, the initial state and the intermediate state may be excluded, and the cumulative score exhibiting the maximum value may be obtained from the cumulative scores related to the final state.

Hereinafter, a specific example of a speech recognition apparatus to which the above-described maximum cumulative score calculation method is applied will be described with reference to FIG. In FIG. 1, an audio signal input from a microphone 1 is A / D converted by an A / D converter 2 and transmitted to an 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 creating unit 4 calculates the symbol output probabilities of all different states constituting the all phoneme HMM from the extracted acoustic parameters and all phoneme HMMs stored in the HMM data memory 9. o is calculated to create a likelihood table. Then,
The reference cumulative score calculation unit 5 for reject determination uses the created likelihood table and the phoneme HMM as shown in FIG. 4 to which the linguistic connection constraint and the phoneme environment connection constraint are added as described above. 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 storage unit for reject determination is updated. On the other hand, the recognition task cumulative score calculation unit 6
Calculates the recognition task cumulative score, which is the cumulative value of the generation probability of the recognition task, for each frame using the likelihood table, each phoneme HMM of the HMM data memory 9, and the recognition task dictionary 12, and stores the recognition task cumulative score. The content of the unit 11 is updated. When the above-described processing for the final frame is completed, the determination unit 7 stores the maximum value of the reject determination reference cumulative score stored in the reject determination reference cumulative score storage unit 10 and the recognition task cumulative score storage unit 11. The recognition result is output or rejected based on the normalized value of the difference between the recognition task cumulative score and the maximum value. The speech recognition control unit 8 controls the A / D converter 2, the acoustic analysis unit 3, the likelihood table creation unit 4, the reference cumulative score calculation unit 5 for reject determination, the recognition task cumulative score calculation unit 6, and the determination unit 7. Then, a voice recognition process is performed.

Here, the phoneme HMM stored in the HMM data memory 9 is a phoneme environment-dependent phoneme HMM, and the phoneme environment connection constraint information and the linguistic information as shown in FIGS. Connection constraint information is added.

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

In step S1, the reference cumulative score for reject determination (hereinafter, simply referred to as reject cumulative score) gr _j (j: state number) stored in the reference cumulative score storage for reject determination 10 is initialized. You. In step S2, the recognition task cumulative score (hereinafter simply referred to as the 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 parameters of the frame i from the digital audio signal.

In step S8, the likelihood table creating unit 4 converts all the phonemes HM stored in the HMM data memory 9 from the extracted acoustic parameters of the frame and the data of the phonemes HMM in the HMM data memory 9.
The symbol output probabilities o _j (i) of all the different states j constituting M are calculated. Here, the phoneme HMM
Are created by independent learning for tasks using phoneme balance words of many speakers. In step S9, the likelihood table creation unit 4 performs normalization and logarithmic conversion on the calculated symbol output probabilities o _j (i) of the respective states to obtain likelihoods of the states. The corresponding frame of the degree 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, and the likelihood table is held in an internal memory or the like of the likelihood table creation unit 4, and the reject cumulative score gr in the frame i is stored. _It may be deleted after the calculation of _j (i) and the cumulative cumulative score gt _k (i) is completed. Thus, the amount of storage required for the voice recognition processing operation is reduced.

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

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

In step S11, the recognition task cumulative score calculation unit 6 causes the recognition task cumulative score storage unit 11
Is read out and the cumulative recognition 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 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 by the calculated recognition cumulative score gt _k (i). Here, the recognition task is represented by a concatenation of phoneme HMMs. Therefore, the recognition cumulative score gt _k of each recognition task k in the previous frame (i-1) (i- 1), a symbol output in the frame i the probability o _k (i) and state transition probability p _k (i) Is accumulated to obtain the cumulative recognition score gt _{k of the} frame i.
(i) is calculated.

In other words, the recognition task cumulative score calculation unit 6 calculates 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 determined whether or not the frame i is the last frame. As a result, if it is not the last 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 last frame, the process proceeds to step S13. In step S13, the frame number i is set in the frame number I. Step S14
Then, the determination unit 7 searches the maximum value among the reject cumulative scores gr _j (I) stored in the reject determination reference cumulative score storage unit 10 and sets the maximum value as the reject determination maximum reference cumulative 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 is set as the recognition task maximum cumulative score Lt.

In step S15, the determination unit 7 determines
The retrieved maximum reference cumulative score L for reject determination
From r and the recognition task maximum cumulative score Lt, a normalized reject determination value L 'is calculated by equation (2). L ′ = (Lr−Lt) / I (2) where the maximum reference cumulative score Lr for reject determination is
And the maximum cumulative score Lt of the recognition task is the number of frames I
It increases in proportion to. Therefore, both cumulative scores Lr,
The difference of Lt is normalized by the number of frames I. In step S16, the determination unit 7 determines whether the calculated value of the normalized rejection determination value L 'is smaller than a threshold value. As a result, if it is smaller than the threshold, the process proceeds to step S18, and if it is not smaller than the threshold, the process proceeds to step S17.
Proceed to. In step S17, the determination unit 7 determines that the utterance content is an unknown utterance, and rejects the phoneme string presenting the recognition task maximum reference cumulative score Lt, thus ending the voice recognition processing operation. In step S18, the determination unit 7 determines that the utterance content is included in the recognized vocabulary, outputs a phoneme string exhibiting the maximum reference cumulative score Lt of the recognition task as a recognition result, and ends the voice recognition processing operation. I do.

As described above, in the present embodiment,
The HMM data memory 9 registers a phoneme environment-dependent phoneme HMM obtained by adding phoneme environment connection constraint information and linguistic connection constraint information to HMM boundaries. Then, the likelihood table creation 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 reference cumulative score calculation unit 5 for reject determination determines that the phoneme H to which the phoneme environment connection constraint and the linguistic connection constraint have been 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 contents of the reject determination reference cumulative score storage unit 10 are updated. On the other hand, the recognition task cumulative score calculation unit 6 stores the phoneme HMM in the likelihood table.
Then, along with the route set according to the recognition task, the recognition cumulative score gt _k (i) is calculated for each recognition task k, and the contents of the recognition task cumulative score storage unit 11 are updated. When the above-described processing is completed up to the final frame, the determination unit 7 stores the two cumulative scores gr _j (i), stored in the reject determination 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 determination of the recognition result is performed using the normalized rejection determination value L ′ obtained based on both the maximum cumulative scores Lr and Lt.

As described above, the reject determination reference cumulative score calculation unit 5 in the present embodiment performs the 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 different states in the immediately preceding frame in the two phoneme HMMs to a shared state j in the frame. (The smaller of the reject cumulative scores) is cut off in the frame by the Viterbi algorithm. As a result, it is not necessary to calculate the reject cumulative score gr _j (i) for the cut-off phoneme HMM after the next frame, and the amount of calculation can be reduced accordingly. On the other hand, in the “likelihood correction method using syllable recognition” described in a paper by Watanabe et al. Described in the section of the prior art, the generation probability is calculated using a syllable network as shown in FIG. It has been calculated. Therefore, even if there is partial state sharing between the syllable HMMs, a cumulative score is obtained independently for each syllable HMM. Therefore, the calculation amount of the cumulative score is not reduced. That is, according to the present embodiment, when the phoneme HMM to be used includes the state-sharing phoneme HMM, the processing amount at the time of reject determination is greatly 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 operation processing can be further reduced by limiting the phoneme HMM to be connected. Furthermore, an impossible route is deleted in advance, and the reject cumulative score gr _j (i) can be accurately calculated, and the reject determination accuracy can be improved.

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 stored in the data memory 9. Therefore, unlike the related art, it is not necessary to store two networks for the reject determination and the recognition task, and the amount of storage can be reduced. In the above embodiment, the reject cumulative score gr _{j is obtained} from the phoneme HMM stored in the HMM data memory 9.
Since (i) and the cumulative recognition score gt _k (i) are calculated, the rejection determination process described above is applicable whether the recognition task is a word or continuous speech recognition using a syntax 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 constraint information and linguistic connection constraint information are added to the HMM boundary as constraint information. However, the present invention is not limited to this, and only the phoneme environment connection restriction information may be added as the restriction information. In addition, there is no problem if a non-phoneme environment-dependent phoneme HMM is registered in the HMM data memory and only linguistic connection constraint information is added to the HMM boundary. However, in that case, the calculation accuracy of the reject cumulative score gr _j (i) becomes low.

[0041]

As is apparent from the above description, the speech recognition apparatus according to the first aspect of the present invention creates a likelihood table by a likelihood table creation unit based on acoustic parameters and HMMs in subword units, and performs rejection determination. The reference cumulative score calculation unit calculates the maximum reference cumulative score along the route set on the likelihood table according to the constraint based on the state transition constraint information of the HMM by the Viterbi algorithm, and the recognition task cumulative score calculation unit Since the maximum cumulative score along the path according to the HMM and each recognition task on the likelihood table is calculated and the reject determination unit determines rejection of the utterance content based on the difference between the two maximum cumulative scores, When calculating the maximum reference cumulative score for judgment, the state of multiple HMMs is 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. Further, the maximum reference cumulative score for the reject determination and the maximum cumulative score for the recognition task are the same subword unit HMM.
, It is not necessary to provide separate networks for reject determination and recognition tasks, and the storage capacity can be reduced.

At this time, since the state transition constraint information of the HMM includes connection constraint information between phonemes in the recognition target language at the HMM boundary, the path set on the likelihood table includes the linguistic connection constraint. Is given. Therefore, according to the present invention, the calculation of the maximum operation amount of the reference cumulative score with there is less, the maximum reference cumulative score for HMM is limited in accordance with the recognition target language reject judgment to be connected for determination Li object Accuracy improves.

A speech recognition apparatus according to a second aspect of the present invention.
The same acoustic analysis unit and HMM
Data memory, likelihood table creation section, reject judgment reference
Cumulative score calculator, recognition task cumulative score calculator, and
Since it has a reject judgment unit,
When calculating large reference cumulative score, multiple H in the same frame
When sharing the state of the MM,
(1) The amount of operation is greatly reduced. As a result, the maximum
The storage amount of the illumination cumulative score can be reduced. In addition,
Maximum reference cumulative score for object judgment and recognition task
The maximum cumulative score of the
Calculated from the likelihood table
Need to provide separate networks for recognition and recognition tasks
Storage capacity can be reduced. Further, the above HMM
The HMM stored in the data memory is a phoneme environment-dependent phoneme HMM, and the state transition constraint information of the HMM includes connection constraint information based on the phoneme environment at the HMM boundary, and is set on the likelihood table. The path is provided with the above-described phoneme environment connection constraint. Therefore, according to the present invention, the calculation amount of the maximum reference cumulative score for determination Li object becomes less, phoneme HMM being linked to calculate the accuracy of the maximum reference cumulative score for rejection determination are specified by the phoneme environment Will be very good.

[Brief description of the 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. 1;

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

FIG. 4 is a diagram illustrating an example of a path when a cumulative score is obtained from a likelihood table by a Viterbi algorithm.

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

FIG. 6 is a diagram illustrating an example of a connection constraint between phonemes following FIG. 5;

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

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

FIG. 9 is an explanatory diagram of an HMM.

[Explanation of symbols]

3 ... Acoustic analysis unit 4 ... Likelihood table creation unit 5 ... Refuse judgment reference cumulative score calculation unit
6: recognition task cumulative score calculation unit, 7: determination unit, 8 ...
9: HMM data memory, 10: reject determination reference cumulative score storage unit, 1
1 ... Recognition task cumulative score storage unit.

Continuation of the front page (56) References JP-A-8-6588 (JP, A) Seiichi Miki, Tatsuya Kawahara, Shuji Doshita, Rejection of utterances deviating from the syntactic constraints of tasks, The Acoustical Society of Japan, 1994 Proceedings of the Fall Meeting of the Japanese Research Conference, Japan, October 1994, 1-Q-1, p. 143-144 Seiichi Nakagawa, Speech Recognition by Probabilistic Model, Japan, IEICE, July 1988, p. 96 Hiroki Ohnishi, Masanori Miyatake, Study of Continuous Speech Recognition Using Phoneme Environment, Proceedings of the Acoustical Society of Japan Fall Meeting, 1990, September 1990, 2-8-4, p. 53-54 Yoshiaki Noda, Shigeki Sagayama, HMM-LR speech recognition by A * beam search using forward likelihood, IEICE technical report [Speech], Japan, June 17, 1994, SP94-23, p. . 1-7 Takao Watanabe and Satoshi Tsukada, Rejection of Unknown Utterance by Likelihood Correction Using Syllable Recognition, IEICE Journal D-II, Japan, December 1992, Vol. J75-D-II No. 12, p. 2002-2009 (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 15/14 G10L 15/18 G10L 15/28 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. Extracting acoustic parameters from input speech
An acoustic analysis unit and an HMM in subword units having state transition constraint information are stored.
The stored HMM data memory, the extracted acoustic parameters and the stored H
Based on the MM, all the states that make up the HMM
Likelihood table for calculating local likelihood and creating a likelihood table
And a constraint based on the HMM state transition constraint information.
A route is set on the likelihood table, and the
Calculate large reference cumulative score by Viterbi algorithm
A reference cumulative score calculation unit for reject determination, and the HMM and each recognition module on the likelihood table.
To calculate the maximum cumulative score along the route
Knowledge task cumulative score calculation unit, and the reject determination reference cumulative score calculation unit
The maximum reference cumulative score and the above-mentioned recognition task cumulative score
Calculate the difference between the maximum cumulative scores calculated by the
If the value is equal to or greater than the predetermined value, the utterance content is unknown
A reject decision unit that decides that it is a story and rejects it
Preparationhand, HMM status stored in the HMM data memory
The transition constraint information is the sound in the target language at the HMM boundary.
Including connection constraint information between elements Voice recognition device characterized by the following:
Place. (2)Extract acoustic parameters from input speech
An acoustic analysis unit, An HMM in subword units having state transition constraint information is stored.
An HMM data memory being loaded; The extracted acoustic parameters and the stored H
Based on the MM, all the states that make up the HMM
Likelihood table for calculating local likelihood and creating a likelihood table
File creation department, According to the constraints based on the state transition constraint information of the above HMM,
A route is set on the likelihood table, and the
Calculate large reference cumulative score by Viterbi algorithm
A reject determination reference cumulative score calculation unit; The HMM and each recognition module on the likelihood table
Along the route To calculate the maximum cumulative score
Knowledge task cumulative score calculation unit, The reject judgment reference cumulative score calculation unit calculates
The maximum reference cumulative score and the above-mentioned recognition task cumulative score
Calculate the difference between the maximum cumulative scores calculated by the
If the value is equal to or greater than the predetermined value, the utterance content is unknown
A reject decision unit that decides that it is a story and rejects
Preparation The HMM stored in the HMM data memory is
The elementary environment-dependent phoneme HMM, and the state of the HMM stored in the HMM data memory.
The transition constraint information is connected by the phoneme environment at the HMM boundary.
A speech recognition device characterized by including about information.