JP3039095B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speech recognition apparatus,
In particular, the present invention relates to rejection of utterances other than a recognition target word, detection of an unknown word, and improvement of word spotting for detecting the presence of a voice belonging to a standard pattern category from a part of input unknown voices.

[0002]

2. Description of the Related Art Conventionally, in speech recognition, as described on pages 149 to 177 of "Digital Speech Processing" published by Tokai University Press (hereinafter referred to as reference [1]).
Recognition has been performed by comparing and collating the input unknown voice with a pre-registered standard pattern to be recognized to obtain a similarity, and selecting a category of the standard pattern that gives the maximum similarity. Here, as the similarity, one based on a distance between feature vectors, one based on a feature vector appearance probability, or the like is used. By using the similarity obtained in this way, when an unknown word that is not registered as a recognition target standard pattern is input, it can be determined as an unknown word. For example, IEICE Technical Report, Vol. 89, N
o. 91, June 1989, page 1-8, "Development of Unspecified Speaker Word Speech Recognition Device for Automatic Ticket Vending Machine under High Noise" (hereinafter referred to as Document [2]). There is a method in which when an obtained similarity is smaller than a predetermined threshold value, the word is determined to be an unknown word and rejected. However, the magnitude of the similarity calculated here is
It varies greatly depending on the speaker and vocal environment. For this reason, when the speaker and the utterance environment are different, it is necessary to set different threshold values for each speaker and each utterance environment in order to increase the accuracy of unknown word detection and rejection. There is a problem that it is necessary. Therefore, Japanese Patent Laid-Open No. 4-2
As described in US Pat. No. 55900 (hereinafter referred to as reference [3]), a similarity between an input unknown voice and a pre-registered standard pattern of a recognition target is obtained, and a restriction on the recognition target is removed. There is a method in which the similarity in the case where the constraint due to language is weakened is obtained as a reference similarity, and the previous similarity is corrected using the reference similarity. The influence of the speaker and the environment appearing in the input speech appears in both cases of obtaining the similarity with the standard pattern to be recognized and the reference similarity, so they were corrected by this method. It can be said that the similarity is offset by the influence of the speaker and the environment. Thus, if the corrected similarity is smaller than a predetermined threshold value and it is determined that the word is unknown, a highly accurate reject can be realized.

[0003] Word spotting for detecting the presence of speech belonging to the standard pattern category from a part of input unknown speech is described in pp. 176-177 of Ref. (Ref. [6]), pp. 87-89, a comparison is made between an input partial section of an unknown voice and a pre-registered standard pattern to obtain a similarity, and the obtained similarity is calculated. When the degree is larger than a predetermined threshold value, there is a method of determining that a voice of the category of the standard pattern exists and performing recognition.

[0004]

Even in word spotting, high-precision speech detection requires a great deal of labor due to the change in the degree of similarity depending on the speaker and the utterance environment. In order to solve this problem, if the above-described similarity correction method is applied and word spotting is performed based on the corrected similarity, it is considered that highly accurate word spotting is possible. However, the above-described similarity correction method cannot be directly applied to a speech recognition method in which endpoints are not fixed, such as word spotting. An object of the present invention is to apply a method of correcting similarity to word spotting so that similarities can be compared on the same scale even when speakers and utterance environments are different, and to perform stable voice detection. .

[0005]

According to a first aspect of the present invention, there is provided a speech recognition apparatus comprising: an analyzing unit for converting an input speech signal into a time series of a frame of a feature vector; and a feature vector constituting a standard pattern registered in advance. And an inter-vector similarity calculation unit that calculates an inter-vector similarity between each of the input audio signals and a feature vector of the input audio signal, and the inter-vector similarity in each frame of the input audio signal. A similarity correction unit that calculates a frame correction similarity corrected using the maximum value of the above, a similarity accumulation unit that accumulates the frame correction similarity to obtain a correction similarity, and identifies a voice based on the correction similarity. It is characterized by having an identification unit that performs the operation.

According to a second aspect of the present invention, there is provided a speech recognition apparatus for converting an input speech signal into a time series of a frame of a feature vector, and converting the feature vector time series and a pre-registered standard pattern into a terminal frame. A comparison / matching unit that determines a standard pattern that gives the maximum similarity as a recognition result and obtains a start frame corresponding to the maximum similarity, and a unit standard pattern storage unit that holds a standard pattern of a recognition unit And a reference similarity calculation for calculating, as a reference similarity, a maximum value of similarity between a standard pattern obtained by combining the standard patterns of the recognition units in a predetermined order and a feature vector time series from the start frame to the end frame. Determining a corrected similarity obtained by correcting the similarity determined by the comparing unit and the comparison matching unit using the reference similarity. And having a similarity correcting unit.

A speech recognition apparatus according to a third aspect of the present invention is characterized in that the speech recognition apparatus has a reject unit that outputs a reject signal when the corrected similarity obtained by the similarity correction unit is smaller than a predetermined threshold.

[0008] A speech recognition apparatus according to a fourth aspect of the present invention is characterized in that the speech recognition apparatus further comprises a speech detection unit for detecting as a speech when the corrected similarity obtained by the similarity correction unit is larger than a predetermined threshold.

According to a fifth aspect of the present invention, there is provided a speech recognition apparatus comprising: an analysis unit for converting an input speech signal into a time series of a frame of a feature vector; a standard pattern storage unit for holding a standard pattern to be recognized; A unit standard pattern storage unit that holds a standard pattern, a reference pattern generation unit that combines the standard patterns of the recognition units in a predetermined order and forms a reference pattern, and combines the reference pattern and the standard pattern of the recognition target. A recognition pattern generation unit to be a recognition pattern, comparison and collation by determining an end frame between the feature vector time series and the recognition pattern, and comparison and collation for obtaining as a recognition result a standard pattern constituting a recognition pattern that gives the maximum similarity And the maximum value of the similarity when the reference pattern and the end frame are determined as the end and compared and compared. A reference similarity calculation unit that obtains the similarity, a similarity correction unit that obtains a corrected similarity obtained by correcting the maximum similarity obtained by the comparison / matching unit using the reference similarity, and a similarity correction unit that obtains the corrected similarity. A voice detection unit that detects the voice as the voice when the corrected similarity is larger than a predetermined threshold.

[0010]

According to the present invention, the similarity between a feature vector time series of an input voice and a standard pattern registered in advance is obtained, and the similarity in the case where the constraint of a word dictionary is removed and the constraint by language is weakened is obtained. When the similarity is obtained and corrected, the similarity is obtained for each frame of the input voice, thereby enabling application to word spotting.

[0011] In the speech recognition apparatus according to the first invention,
Consider a case where speech is recognized using a standard pattern registered in advance. Here, each standard pattern is constituted by a sequence of feature vectors.

First, an input speech signal is converted into a time series of feature vector frames by an analysis unit. For the analysis here, a method based on mel-cepstrum, a method based on LPC analysis, etc., shown on pages 32-98 of document [1] can be used.

Next, in the inter-vector similarity calculation unit,
The similarity between the feature vector obtained by the analysis unit and the feature vector constituting the standard pattern registered in advance is calculated for each frame of the input audio signal. As a method for obtaining the similarity between vectors, there is a method described in reference [1].
A method based on the distance between vectors and a hidden Markov model (hereinafter referred to as HMM
) Can be used. For HMM, S.M. E. FIG. Levinson (SE Levinson), L. R. LR Rabiner; M. Sondhi et al., The Bell System Te.
chemical Journal), Vol. 62, N
o. 4, April 1983, pp. 1035-1074, entitled "An Introduction to the Application of the Theory of Probabilistic Functions of a Markov."
Process Introduction to Automatic Speech Recognition (An Introduction to the
Application of the Theor
y of Probabilistic Function
s of a Markov Process to
Automatic Speech Recognit
ion) "(hereinafter referred to as Document 4).

The similarity correction unit calculates the maximum value of the inter-vector similarity for each frame, corrects all the inter-vector similarities of the frame using the obtained maximum value, and calculates the frame correction similarity. Is done. As the correction method here, a method of obtaining a value based on the difference between the inter-vector similarity and the maximum value of the inter-vector similarity, or a value based on the ratio of the inter-vector similarity and the maximum value of the inter-vector similarity is used. The required method can be used.

In the similarity accumulating section, a corrected similarity obtained by accumulating the frame corrected similarity with respect to the input audio signal is obtained. As a method of accumulating the frame correction similarity, in the case of comparison and collation with a word pattern, the correction similarity is calculated based on the dynamic programming method, as shown on pages 154 to 165 of document [1]. How to accumulate, literature [4]
, A method of accumulating the corrected similarity based on the HMM can be used.

The identification section identifies the speech based on the corrected similarity and obtains a recognition result. Here, as the method of identifying the voice, a method of using the one having the largest correction similarity as the recognition result can be used.

The influence of the speaker and the environment appearing in the input voice appears in both cases of obtaining the similarity and the maximum value of the similarity used for the correction. In the similarity, the influence of the speaker and the environment is offset.

Reference [3] also describes a method based on the maximum value of the inter-vector similarity for each frame. This is because the calculation of the similarity with the standard pattern obtained by accumulating the inter-vector similarity for each frame and the calculation of the reference similarity obtained by accumulating the maximum value of the inter-vector similarity for each frame are separately performed. After this, the similarity with the standard pattern is corrected by the reference similarity. The section for obtaining the similarity with the standard pattern and the section for obtaining the reference similarity need to be the same section. However, when performing word spotting, since the end point is not determined in advance, reference [3] In the method [1], it is necessary to calculate the reference similarity for the same section after calculating the similarity with the standard pattern. For this reason, the amount of calculation after obtaining the similarity with the standard pattern increases, and the method is not suitable for word spotting.

On the other hand, in the speech recognition apparatus according to the first aspect of the invention, the corrected similarity for each frame is accumulated. The section for obtaining and the section for obtaining the similarity used for the correction always coincide with each other. Therefore, there is no need to calculate the reference similarity after performing the cumulative calculation of the similarity for each frame.

From these facts, the speech recognition apparatus according to the first aspect of the present invention suppresses the difference in the magnitude of similarity due to the difference in speaker and environment, and the similarity suitable for application to word spotting. This is a correction method.

As described above, according to the first aspect, it is possible to obtain a corrected similarity that is not affected by the speaker or the environment.

In the speech recognition apparatus according to the second aspect of the present invention, the similarity between the time series of the feature vector frame obtained by the analysis unit and the plurality of standard patterns registered in advance is determined by the comparison and collation unit. , And a standard pattern that gives the maximum similarity is obtained as a recognition result, and a start frame corresponding to the maximum similarity is obtained. Here, as a method of comparison and matching with a word pattern, as shown on pages 154 to 165 of document [1], a feature vector time series is held as a standard pattern, and similarity is determined based on the distance between feature vectors. There are a method of performing comparison and comparison based on a dynamic programming method, and a method of performing comparison and comparison based on an HMM as described in reference [4]. The method of finding the start frame corresponding to the maximum similarity is described in IEICE Transactions,
Vol. J71-D, No. As described in “Speeding up DP Matching by Integrating Frame Synchronization, Beam Search, and Vector Quantization” on page 9, pages 1650-1659 (hereinafter referred to as reference [5]), In selecting the collation path, a method of sequentially transmitting the start frame and determining the maximum similarity and simultaneously obtaining the start frame can be used.

In the reference similarity calculation unit, the feature vector time series of the input speech in the range from the start frame to the end frame obtained by the comparison / matching unit and the recognition unit standard pattern stored in the unit standard pattern storage unit are stored. The similarity is obtained by comparing and collating with a plurality of standard patterns combined in a predetermined order, and the maximum value of the similarity is set as the reference similarity. As a method of comparing and recognizing a recognition unit pattern with a pattern combined in a predetermined order, as described in the document [5], the predetermined order is described in advance in the form of a network. A method of accumulating corrected similarities based on continuous speech recognition by simple frame-synchronous DP matching, and the method of “Speech recognition using a stochastic model” published by the Institute of Electronics, Information and Communication Engineers (hereinafter referred to as reference [6]).
A method of accumulating the corrected similarity based on a continuous speech recognition algorithm by the HMM as shown on page 0-50 can be used. Further, as the recognition unit standard pattern here, syllables, phonemes, words, and the like as shown on page 5 of document [1] can be used.

Next, the similarity correction unit corrects the similarity with the standard pattern obtained by the comparison / matching unit using the reference similarity, and calculates the corrected similarity. Examples of the correction method include a method of calculating a value based on the difference between the similarity with the standard pattern and the reference similarity, and a method of calculating a value based on the ratio of the similarity with the standard pattern and the reference similarity. Can be used.

The influence of the speaker and the environment appearing in the input voice appears in both cases of obtaining the similarity with the standard pattern and obtaining the reference similarity used for correction. In the corrected similarity obtained in this way, the influence of the speaker and the environment is offset.

Even if the end point is not fixed in advance, the method of calculating the similarity to the standard pattern, then obtaining the end point, calculating the likelihood for reference, and correcting it is used for word spotting. Can be applied.

As described above, also in the speech recognition apparatus according to the second invention, it is possible to suppress the difference in the magnitude of the similarity due to the difference in the speaker and the environment.

In the third aspect, a reject signal is generated when the correction similarity is smaller than a predetermined threshold in the reject unit. Here, the difference in the magnitude of the similarity due to the difference in the speaker and the environment is corrected, and the corrected similarity can be rejected using a certain threshold.

In the document [3], it is necessary to perform the correction based on the reference similarity after calculating the similarity with the standard pattern. Since the corrected similarity is calculated, the corrected similarity is calculated by the same calculation as the similarity calculation with the standard pattern in the document [3], and there is an advantage that it is not necessary to perform correction later.

In the fourth invention, a case where the corrected similarity is larger than a predetermined threshold value is detected as a sound by the sound detection unit. Here, in the corrected similarity, the difference in the magnitude of the similarity due to the difference in the speaker and the environment is corrected, and the voice can be detected using a certain threshold.

In the fifth aspect, the reference pattern generation unit combines the standard patterns of the recognition units stored in the unit standard pattern storage unit in a predetermined order to generate a reference pattern. Further, as the recognition unit standard pattern here, syllables, phonemes, words, and the like as shown on page 5 of document [1] can be used.

The recognition pattern generation section generates a recognition pattern obtained by combining the reference pattern and the standard pattern.

The comparison / matching section determines the end frame of the feature vector time series and the recognition pattern, compares and matches them, and obtains, as a recognition result, a standard pattern constituting a recognition pattern giving the maximum similarity. Here, as a method of comparison and matching with a word pattern, as shown on pages 154 to 165 of document [1], a feature vector time series is held as a standard pattern, and similarity is determined based on the distance between feature vectors. How to calculate and compare and match based on dynamic programming,
There is a method of comparing and matching based on the HMM as described in reference [4]. Further, as a method of comparison and comparison with a reference pattern, continuous speech recognition by frame-synchronous DP matching as described in reference [5],
H as shown on pages 40-50 of document [6]
A continuous speech recognition algorithm based on MM can be used.

The reference similarity calculation unit calculates the similarity by comparing and collating the feature vector time series with the reference pattern by defining the end frame, and calculates the maximum similarity as the reference similarity. As a method of comparison and comparison with a reference pattern, continuous speech recognition by frame-synchronous DP matching as described in reference [5] or HMM as described in page 40-50 of reference [6] is used. A continuous speech recognition algorithm can be used.

Next, the similarity correction unit obtains a corrected similarity obtained by correcting the maximum similarity obtained by the comparison / matching unit using the reference similarity. As a correction method here,
A method of obtaining a value based on the difference between the similarity with the standard pattern and the reference similarity, a method of obtaining a value based on the ratio of the similarity with the standard pattern and the reference similarity, or the like can be used.

Next, when the corrected similarity obtained by the similarity correction unit is larger than a predetermined threshold value, the voice detection unit detects it as voice.

The influence of the speaker and the environment appearing in the input voice appears in both cases of obtaining the similarity with the standard pattern and obtaining the reference similarity used for correction. In the corrected similarity obtained in this way, the influence of the speaker and the environment is offset.

In the second invention, the similarity between the feature vector time series and a plurality of standard patterns registered in advance is calculated by determining the end frame, and the start frame corresponding to the maximum similarity is obtained. Then, the part between the start frame and the end frame is extracted from the feature vector time series,
Calculation of reference similarity is performed. Since the reference similarity is redone for each end frame, there is a problem that the amount of calculation increases. In the fifth aspect, the similarity based on the comparison and comparison with the recognition pattern obtained by combining the reference pattern and the standard pattern is corrected using the reference similarity based on the comparison and comparison with the reference pattern. The result of the comparison and comparison between the input voice and the recognition pattern is divided into a portion corresponding to the standard pattern to be recognized and a portion corresponding to the reference pattern. This means that word spotting for associating the partial section with the standard pattern has been performed. It is considered that substantially the same reference pattern is associated with the portion corresponding to the reference pattern in the calculation of the similarity for reference, and substantially the same value is obtained for the similarity. Therefore, by correcting the similarity with the recognition pattern by the similarity with the reference pattern, the similarity with the standard pattern was corrected using the similarity with the reference pattern in the section corresponding to the standard pattern. This means that the corrected similarity has been obtained. As a result, it is possible to perform word spotting adapted to correction of similarity by comparison and collation with a fixed starting end, and to perform high-precision word spotting with a small amount of calculation.

[0039]

Embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of one embodiment of the first, third and fourth inventions.

This speech recognition apparatus includes an analysis unit 1, an inter-vector similarity calculation unit 2, a similarity correction unit 3, a similarity accumulation unit 4,
An identification unit 5, a reject unit 6, and a voice detection unit 7 are provided.

The analysis unit 1 analyzes the characteristics of the input audio signal I and converts it into a characteristic vector time series V.

The inter-vector similarity calculation unit 2 calculates and outputs an inter-vector similarity D between a feature vector of one frame of the feature vector time series V and a feature vector constituting a standard pattern registered in advance. is there.

The similarity correction unit 3 receives the inter-vector similarity D, calculates the maximum value of the inter-vector similarity for one frame, and corrects the maximum value of the inter-vector similarity for each frame with the obtained maximum value. The correction similarity F is output.

The similarity accumulator 4 calculates the frame correction similarity F
Is output as the accumulated corrected similarity C.

The identification section 5 identifies a voice using the corrected similarity C and outputs the result as a recognition result A.

The reject unit 6 outputs a reject signal J when the correction similarity C is smaller than a predetermined threshold.

The voice detector 7 detects the recognition result A when the correction similarity C is larger than a predetermined threshold.

Next, the operation of the embodiment shown in FIG. 1 will be described.

The input speech signal I is input to the analysis unit 1 and is converted into a feature vector time series V by feature analysis. For the analysis here, for example, a method based on mel-cepstral as shown on page 155 of document [1] can be used.

The feature vector time series V is input to the inter-vector similarity calculation unit 2 and the similarity with the feature vectors constituting a plurality of standard patterns registered in advance is calculated. Become. As a method of calculating the inter-vector similarity, for example, 154-
A method based on the inter-vector distance as described on page 161 can be used.

The inter-vector similarity D is input to the similarity correction unit 3 to determine the maximum value of the inter-vector similarity for one frame, and to correct using the maximum value of the inter-vector similarity for each frame. And the frame correction similarity F
Is output as As a correction method, a method of obtaining a value based on the difference between the inter-vector similarity and the maximum value of the inter-vector similarity can be used.

The frame correction similarity F is calculated by the similarity accumulating unit 4
And outputs a corrected similarity C obtained by accumulating the frame corrected similarities. As a method of accumulating the frame correction similarities, for example, a method of accumulating the correction similarities based on the dynamic programming, as shown on pages 154 to 165 of document [1], can be used.

The corrected similarity C is input to the identification unit 5, and the voice is identified using the corrected similarity.
Is output as As a method of identification, for example, a method of selecting a standard pattern having the maximum corrected similarity can be used.

As described above, according to the first aspect, it is possible to obtain a corrected similarity that is not affected by the speaker or the environment.

According to the third aspect, when the correction similarity C is smaller than the predetermined threshold value, the rejection unit J outputs the rejection signal J.

According to the fourth aspect, when the corrected similarity C is larger than a predetermined threshold, the voice detection unit 8 detects the voice as voice and outputs the recognition result A.

FIG. 2 is a block diagram of one embodiment of the second and fourth inventions.

This speech recognition apparatus includes an analysis unit 11, a comparison / collation unit 12, a unit standard pattern storage unit 13, a reference similarity calculation unit 14, a similarity correction unit 15, and a speech detection unit 17.

The analysis section 11 analyzes the characteristics of the input audio signal I and converts it into a characteristic vector time series V.

The comparison / collation unit 12 calculates the characteristic vector time series V
And a standard pattern registered in advance to determine the end frame E and compare and collate the standard frame. The standard pattern giving the maximum similarity S is obtained as the recognition result A, and the start frame T corresponding to the maximum similarity is obtained. is there.

The unit standard pattern storage unit 13 stores a unit standard pattern.

The reference similarity calculation unit 14 calculates a standard pattern obtained by combining unit standard patterns registered in the unit standard pattern storage unit 13 in a predetermined order and a feature vector time series V from a starting frame T. The feature vector time series obtained by extracting the section up to the end frame E is compared and collated, the similarity is calculated for each standard pattern, and the maximum similarity is obtained as the reference similarity R.

The similarity correction unit 15 corrects the similarity S using the reference similarity R to obtain a corrected similarity C.

The voice detecting section 7 detects the recognition result A when the corrected similarity C is larger than a predetermined threshold value.

Next, the operation of the embodiment shown in FIG. 2 will be described.

The input voice signal I is input to the analysis unit 11 and is converted into a feature vector time series V by feature analysis.

The feature vector time series V is compared with the comparison /
The similarity is calculated by comparing and comparing a plurality of standard patterns registered in advance with the end frame E, and the standard pattern giving the maximum similarity S and the maximum similarity is obtained as the recognition result A. And the starting frame T corresponding to the maximum similarity is obtained. Here, as a method of calculating the similarity, a method based on DP matching as described in References [1] and [5] and an HMM as described in References [4] and [6] are used. A method can be used.

The feature vector time series V is inputted to the reference similarity calculation unit 14 and is stored in the recognition unit standard pattern storage unit 13 with a portion extracted between the start frame T and the end frame E from the feature vector time series. A plurality of standard patterns obtained by combining the unit standard patterns described above in a predetermined order are compared and collated, and a reference similarity R is output. As the recognition unit standard pattern, a syllable as shown on page 5 of document [1] can be used.

Next, the similarity S and the reference similarity R are input to the similarity correction unit 15, and the corrected similarity C obtained by correcting the similarity S using the reference similarity R is output. As a correction method here, a method of obtaining a value based on the difference between S and R can be used.

As described above, according to the second aspect, it is possible to obtain a corrected similarity that is not affected by the speaker or the environment.

According to the fourth aspect, when the corrected similarity C is larger than a predetermined threshold, the voice detection unit 18 detects the voice as voice and outputs a recognition result A.

FIG. 3 is a block diagram of one embodiment of the fifth invention.

This speech recognition apparatus includes an analysis unit 21, a comparison / collation unit 22, a unit standard pattern storage unit 23, a reference similarity calculation unit 24, a similarity correction unit 25, a speech detection unit 27, a standard pattern storage unit 28, a recognition unit A pattern generator 29 and a reference pattern generator 30 are provided.

The analysis section 21 analyzes the characteristics of the input audio signal I and converts it into a characteristic vector time series V.

The standard pattern storage section 28 stores a standard pattern.

The unit standard pattern storage unit 23 stores unit standard patterns.

The reference pattern generator 30 generates a reference pattern Q in which unit standard patterns stored in the unit standard pattern section are combined in a predetermined order. The recognition pattern generation unit 29 combines the reference pattern Q and the standard pattern stored in the standard pattern storage unit 28 to generate a recognition pattern P.

The comparison / collation unit 22 calculates the characteristic vector time series V
Is compared with the recognition pattern P, and the standard pattern forming the recognition pattern that gives the maximum similarity S is recognized as the recognition result A.
Is what you want.

The reference similarity calculator 24 calculates the reference pattern Q
Is compared with the feature vector time series V, the similarity is calculated for each reference pattern, and the maximum similarity is obtained as the reference similarity R.

The similarity correction unit 25 corrects the similarity S using the reference similarity R to obtain a corrected similarity C.

When the corrected similarity C is larger than a predetermined threshold value, the voice detection unit 27
Is detected.

Next, the operation of the embodiment shown in FIG. 3 will be described.

The input speech signal I is input to the analysis unit 21 and is converted into a feature vector time series V by feature analysis.

The feature vector time series V is obtained by comparing the
The similarity calculated by comparing with the recognition pattern P is calculated, and the standard pattern forming the recognition pattern giving the maximum similarity S and the maximum similarity is obtained as the recognition result A. As a method based on DP matching as shown in References [1] and [5], and H as shown in References [4] and [6]
A method based on MM can be used.

The feature vector time series V is input to the reference similarity calculation unit 24, where it is compared with a plurality of reference patterns Q, and the reference similarity R is output.

Next, the similarity S and the reference similarity R are input to the similarity correction unit 25, and the corrected similarity C obtained by correcting the similarity S using the reference similarity R is output. As the correction method here, for example, a method of obtaining a value based on the difference between S and R can be used. Next, when the corrected similarity C is larger than a predetermined threshold, the voice detection unit 28 detects the voice as voice and outputs a recognition result A.

As described above, according to the fifth aspect, word spotting can be performed by obtaining a corrected similarity that is not affected by a speaker or an environment.

[0089]

As described above, according to the present invention, the similarity between the feature vector time series of the input voice and the registered standard pattern is determined by using the maximum value of the inter-vector similarity of each frame or the input voice and the reference voice. By performing correction using the reference similarity with the standard pattern, the similarity can be compared on the same scale even when the speaker and the utterance environment are different, and stable rejection and word spotting can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the first, third and fourth inventions.

FIG. 2 is a block diagram of one embodiment of the second and fourth inventions.

FIG. 3 is a block diagram of one embodiment of the fifth invention.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 analysis unit 2 inter-vector similarity calculation unit 3 similarity correction unit 4 similarity accumulation unit 5 identification unit 6 reject unit 7 voice detection unit 11 analysis unit 12 comparison and collation unit 13 unit standard pattern storage unit 14 reference similarity calculation unit 15 Similarity correction unit 17 Voice detection unit 21 Analysis unit 22 Comparison and comparison unit 23 Unit standard pattern storage unit 24 Reference similarity calculation unit 25 Similarity correction unit 27 Voice detection unit 28 Standard pattern storage unit 29 Recognition pattern generation unit 30 Reference pattern generation Department

Continuation of the front page (56) References JP-A-4-369696 (JP, A) JP-A-63-56700 (JP, A) JP-A-4-255900 (JP, A) JP-A-4-188200 (JP) , A) JP-A-1-100600 (JP, A) JP-A-5-19786 (JP, A) Patent 2808906 (JP, B2) Proceedings of the Acoustical Society of Japan, Spring Meeting, 1991, I-3- P.28 “Likelihood correction of speech recognition for unknown word detection / rejection” p. 203-204 (accepted by the National Diet Library on August 5, 1991) Proceedings of the Acoustical Society of Japan, Spring Meeting 2004, I-P1-P-11 "Evaluation of adaptive likelihood correction for speech recognition" p. 125-126 (issued on March 17, 1992) (58) Fields surveyed (Int. Cl. ⁷ , DB name) G10L 15/10 G10L 15/08 G10L 15/28 JICST file (JOIS)

Claims (5)

(57) [Claims] An analysis unit configured to convert an input audio signal into a time series of a frame of a feature vector; and a vector between a feature vector constituting a standard pattern registered in advance and a feature vector of the input audio signal. An inter-vector similarity calculator for obtaining a similarity for each frame of the input audio signal, and obtaining a frame correction similarity corrected using the maximum value of the inter-vector similarity in each frame of the input audio signal A speech recognition device, comprising: a similarity correction unit; a similarity accumulation unit that accumulates the frame correction similarity to obtain a correction similarity; and an identification unit that identifies speech based on the correction similarity. 2. The method according to claim 1, wherein the input speech signal is
An analysis unit that converts the feature vector into a time series;
The turn and the end frame are determined and compared and compared.
When a standard pattern that gives similarity is obtained as a recognition result
First, find the starting frame corresponding to the maximum similarity
Comparison / collation unit and unit standard pattern notation that holds the standard pattern of recognition units
Storage unit and a standard pattern of the recognition unit in a predetermined order.
From the standard pattern and the starting frame
The maximum similarity with the feature vector time series up to the end frame
A reference similarity calculator for determining a maximum value as a reference similarity;,  The similarity determined by the comparison / matching unit is referred to by the reference class.
SimilarityToSimilarity correction unit that calculates the corrected similarity corrected by using
And a speech recognition device having: 3. The speech recognition apparatus according to claim 1, further comprising a reject unit that outputs a reject signal when the correction similarity obtained by said similarity correction unit is smaller than a predetermined threshold. 4. The speech recognition device according to claim 1, further comprising a speech detection unit that detects as a speech when the corrected similarity obtained by the similarity correction unit is larger than a predetermined threshold. 5. The method according to claim 1, further comprising:
Analysis unit that converts the time series of the game, and standard pattern storage unit that stores the standard patterns to be recognized
And the unit standard pattern notation that holds the standard pattern of the recognition unit.
Storage unit and a standard pattern of the recognition unit in a predetermined order.
A reference pattern generation unit that combines the reference patterns into a reference pattern, and combines the reference pattern and the standard pattern to be recognized.
A recognition pattern generating unit that sets the feature vector time series and the recognition pattern as end patterns;
Recognition that gives the maximum similarity by defining and comparing frames
Find the standard pattern that constitutes the pattern as the recognition result
A comparison / matching unit that determines the end of the reference pattern and the end frame as an end.
The maximum value of the similarity in the case of comparison is determined as the reference similarity.
A reference similarity calculator that calculates the maximum similarity obtained by the comparison / matching unit.
Similarity to find corrected similarity corrected using reference similarity
Compensator and,  The corrected similarity obtained by the similarity correction unit is
Is detected as voice if it is larger than a predetermined threshold
A speech recognition device having a speech detection unit.