JPH05210396A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To stably detect voices by applying a method for correcting similarity to word spotting so as to compare the similarity in the same scale regardless of difference in speakers or environments. CONSTITUTION:An inter-vector similarity calculation part 2 calculates inter- vector similarity D for each frame between a feature vector V of an input voice I and a standard pattern registered in advance. Next, a similarity correction part 3 corrects the inter-vector similarity D by using the maximum value of the D in the same frame, and frame corrected similarity F is calculated. A similarity accumulation part 4 accumulates the frame corrected similarity F and calculates corrected similarity C. An identification part 5 calculates the standard pattern applying the maximum value of the corrected similarity C as a recognized result A.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a voice recognition device,
In particular, the present invention relates to improvement of word spotting for rejecting utterances other than recognition target words, detecting unknown words, and detecting the presence of speech belonging to a standard pattern category from a part of input unknown speech.

[0002]

2. Description of the Related Art Conventionally, in speech recognition, as described on pages 149-177 of "Digital Speech Processing" published by Tokai University Press (hereinafter referred to as reference [1]).
The recognition was performed by comparing and collating the input unknown voice and the standard pattern of the recognition target registered in advance to obtain the degree of similarity, and selecting the category of the standard pattern that gives the maximum degree of similarity. Here, as the degree of similarity, one based on the distance between feature vectors, one based on the probability of feature vector appearance, or the like is used. By using the similarity thus obtained, when an unknown word that is not registered as a standard pattern to be recognized is input, it can be determined as an unknown word. For example, Technical Report of IEICE, Vol. 89, N
o. 91, June 1989, page 1-8, as described in "Development of an unspecified speaker word speech recognition device for automatic ticket vending machines under high noise" (hereinafter referred to as reference [2]). There is a method of rejecting an unknown word when the obtained similarity is smaller than a predetermined threshold value. 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 thresholds for each speaker and each utterance environment in order to improve the accuracy of unknown word detection and rejection. There is a problem that it is necessary. Therefore, Japanese Patent Application No. 3-6
As described in 0786 (hereinafter referred to as document [3]), the similarity between the input unknown voice and the standard pattern of the recognition target registered in advance is obtained, and the restriction of the recognition target is eliminated, and the language There is a method in which the similarity when the constraint due to is weakened is obtained as the reference similarity, and the previous similarity is corrected using the reference similarity. The influence of the speaker and the environment appearing in the input voice is irrespective of whether the similarity with the standard pattern to be recognized is obtained or the reference similarity is obtained.
Since it appears in both cases as well, it can be said that the similarity corrected by this method cancels out the influence of the speaker and the environment. By this, if the corrected similarity is smaller than a predetermined threshold value, if it is determined as an unknown word,
A highly accurate reject can be realized.

Word spotting for detecting the presence of a voice belonging to the category of the standard pattern from a part of the input unknown voice is described on pages 176-177 of document [1] and pages 87-89 of document [6]. As described above, the similarity is calculated by comparing and collating the input partial segment of the unknown voice with the standard pattern registered in advance, and when the obtained similarity is larger than a predetermined threshold value, the standard pattern There is a method of recognizing by determining that the voice of the category exists.

[0004]

Even in word spotting, a great deal of labor is required for highly accurate voice detection because of the change in the degree of similarity depending on the speaker and the utterance environment. In order to solve this problem, if the above-mentioned 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 voice recognition method such as word spotting in which the end points are not fixed. An object of the present invention is to apply a method of correcting the similarity to the word spotting so that the similarity can be compared on the same scale even when the speaker and the utterance environment 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, which comprises an analysis unit for converting an input speech signal into a time series of feature vector frames, and a feature vector forming a pre-registered standard pattern. And a vector-to-vector similarity calculation unit that obtains the vector-to-vector similarity between the feature vector of the input audio signal for each frame of the input audio signal, and the inter-vector similarity in each frame of the input audio signal , A similarity correction unit that obtains a frame-corrected similarity corrected using the maximum value, a similarity accumulation unit that accumulates the frame-corrected similarity as a corrected similarity, and identifies a voice based on the corrected similarity. It is characterized by having an identifying section that

The speech recognition apparatus of the second invention uses an analysis unit for converting an input speech signal into a time series of feature vector frames, and an end frame for the feature vector time series and a pre-registered standard pattern. A comparison and matching unit that determines and compares and determines a standard pattern that gives the maximum similarity as a recognition result and a starting frame that corresponds to the maximum similarity, and a unit standard pattern storage unit that holds the standard pattern of the recognition unit And a reference similarity calculation for obtaining the maximum value of the similarity between the standard pattern in which the standard patterns of the recognition units are combined in a predetermined order and the feature vector time series from the start frame to the end frame as the reference similarity. Section and the comparison and collation section, the similarity degree is corrected using the reference similarity degree to obtain a corrected similarity degree. And having a similarity correcting unit.

The speech recognition apparatus of the third invention is characterized in that it has a reject unit for outputting a reject signal when the corrected similarity obtained by the similarity correction unit is smaller than a predetermined threshold value.

The voice recognition apparatus of the fourth invention is characterized by having a voice detection unit for detecting as voice when the corrected similarity obtained by the similarity correction unit is larger than a predetermined threshold value.

According to a fifth aspect of the speech recognition apparatus, an analyzing section for converting an inputted speech signal into a time series of feature vector frames, a standard pattern storing section for holding a standard pattern to be recognized, and a recognition unit 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 to form a reference pattern, and combines the reference pattern and the standard pattern to be recognized. A comparison pattern matching unit that recognizes a recognition pattern as a recognition pattern, compares and collates the feature vector time series and the recognition pattern by defining an end frame, and obtains a standard pattern that constitutes a recognition pattern that gives the maximum similarity as a recognition result. Section, refer to the maximum value of the similarity when the reference pattern and the end frame are defined as the end and compared and collated. It is obtained by the reference similarity calculation unit that obtains the similarity, and the similarity correction unit and the similarity correction unit that obtain the corrected similarity by correcting the maximum similarity obtained by the comparison and matching unit using the reference similarity. It is characterized by having a voice detection unit that detects as voice when the obtained corrected similarity is larger than a predetermined threshold value.

[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 the language is weakened by eliminating the constraint of the word dictionary. In the case of finding and correcting the previous similarity, it is possible to apply to word spotting by finding the similarity for each frame of the input voice.

In the speech recognition apparatus according to the first invention,
Consider a case where a voice is recognized using a standard pattern registered in advance. Here, each standard pattern is composed of a series of feature vectors.

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

Next, in the vector similarity calculation unit,
The similarity between the feature vector obtained by the analysis unit and the feature vector forming the standard pattern registered in advance is calculated for each frame of the input voice signal. As a method of obtaining the similarity between vectors, the method of 1 in Reference [1] is used.
Methods based on distances between vectors as described on pages 54-161 and hidden Markov models (hereinafter referred to as HMM
Method) based on the vector appearance probability based on For HMM, see S. E. Levinson and L.E. R. LR Rabiner, and M.R. Bell Systems Technical Journal of MM Sondi et al.
chnical Journal), Vol. 62, N
o. 4, April 1983, pp. 1035-1074, "An Introduction to the Application of the Theory of Probabilistic Functions of A Markov Process to Automatic Speech Recognition".
he Application of the The
ory of Probabilistic Funct
ions of a Marcov Process
to Automatic Speech Recog
Nation) ”(hereinafter referred to as Document 3).

In the similarity correction section, the maximum value of the inter-vector similarity is calculated for each frame, all the inter-vector similarities of the frame are corrected using the calculated maximum value, and the frame correction similarity is calculated. To be done. As the correction method here, a method based on the difference between the vector-to-vector similarity and the maximum value of the vector-to-vector similarity or a value based on the ratio of the vector-to-vector similarity and the vector-to-vector maximum value is used. The method of obtaining can be used.

In the similarity accumulating unit, the corrected similarity is calculated by accumulating the frame correction similarity for the input audio signal. As a method of accumulating frame-corrected similarity here, in the case of comparison and matching with a word pattern, the corrected similarity is calculated based on the dynamic programming method as shown on page 154-165 of document [1]. Accumulation method, reference [4]
A method of accumulating the corrected similarity based on the HMM can be used, as described in.

The discrimination section discriminates the voice based on the corrected similarity and obtains the recognition result. As a method of identifying the voice here, a method of using the one having the largest correction similarity as the recognition result or the like can be used.

The influence of the speaker or the environment appearing in the input voice appears in both the maximum value of the similarity used for the correction when the similarity is calculated, and therefore, the correction thus obtained. In the degree of similarity, the influence of the speaker and the environment is offset.

Document [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 performed separately. After this is performed, 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 word spotting is performed, the end points are not determined in advance, so that the reference [3 In the above method, it is necessary to calculate the reference similarity for the same section after obtaining the similarity with the standard pattern. For this reason, the amount of calculation after obtaining the degree of similarity with the standard pattern is large, which is not suitable for word spotting.

On the other hand, in the speech recognition apparatus according to the first aspect of the present invention, the correction similarity for each frame is accumulated, and the correction similarity thus obtained is similar to the standard pattern. The section for obtaining and the section for obtaining the similarity used for correction always match. Therefore, it is not necessary to calculate the reference similarity after performing the cumulative calculation of the similarity for each frame.

From the above, the voice recognition apparatus according to the first invention suppresses the difference in the degree of similarity due to the difference in the speaker and the environment, and the similarity degree suitable for the application to word spotting is reduced. It is a correction method.

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

In the speech recognition apparatus according to the second aspect of the invention, in the comparison and collation unit, the similarity between the time series of the feature vector frames obtained by the analysis unit and a plurality of standard patterns registered in advance is the end frame. Is calculated and the standard pattern giving the maximum similarity is obtained as a recognition result, and the start frame corresponding to the maximum similarity is obtained. Here, as a method of comparison and collation with a word pattern, as shown on page 154-165 of the document [1], a feature vector time series is held as a standard pattern and the similarity is calculated based on the distance between feature vectors. There is a method of performing comparison and comparison based on dynamic programming, and a method of comparison and comparison based on HMM as described in the document [4]. The method of obtaining the starting frame corresponding to the maximum similarity is as follows:
Vol. J71-D, No. As shown in “Speeding up DP matching by integrating frame synchronization, beam search, and vector quantization” (hereinafter referred to as document [5]), published on pages 9, 1650 to 1659, at the time of comparison and matching. When selecting the matching path, the starting frame can be sequentially transmitted, and the starting frame can be obtained at the same time when the maximum similarity is determined.

In the reference similarity calculation unit, there is a feature vector time series of the input voice in the range from the start frame to the end frame obtained by the comparison and collation unit and the recognition unit standard pattern stored in the unit standard pattern storage unit. 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 collating the recognition unit patterns with the patterns combined in the defined order, by describing the defined order in the form of a network in advance, as shown in document [5]. Method of accumulating the correction similarity based on continuous speech recognition by various frame-synchronous DP matching, and "Speech recognition by probabilistic model" (hereinafter referred to as reference [6]) published by the Institute of Electronics, Information and Communication Engineers.
A method of accumulating the corrected similarities based on the continuous speech recognition algorithm by HMM as shown on page 0-50 can be used. Further, as the recognition unit standard pattern here, a syllable, a phoneme, a word or the like as shown on page 5 of the document [1] can be used.

Next, in the similarity correction unit, the similarity with the standard pattern obtained by the comparison and collation unit is corrected using the reference similarity, and the corrected similarity is calculated. As the correction method here, there are 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 between the similarity with the standard pattern and the reference similarity, and the like. Can be used.

The influence of the speaker and the environment appearing in the input voice appears in both the similarity between the standard pattern and the reference similarity used for correction. In the corrected similarity thus obtained, the influences of the speaker and the environment are offset.

Even if the end points are not fixed in advance, the method of calculating the similarity with the standard pattern, then calculating the reference likelihood by calculating the end points, and performing the correction is used. Can be applied.

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

In the third invention, the reject signal is generated when the corrected similarity is smaller than a predetermined threshold value in the reject unit. Here, the corrected similarity has been corrected for differences in the degree of similarity due to differences in speakers and environments, and can be rejected using a certain threshold value.

In the reference [3], it was necessary to perform the correction based on the reference similarity after obtaining the similarity with the standard pattern, but in the third invention, in each frame according to the first invention. Since the corrected similarity is obtained, it is an advantage that the corrected similarity is obtained by the same calculation as the similarity calculation with the standard pattern in the document [3], and there is no need to make correction later.

According to the fourth aspect of the present invention, the voice detection unit detects a voice when the corrected similarity is larger than a predetermined threshold value. Here, in the corrected similarity, the difference in the degree of similarity due to the difference in the speaker or the environment is corrected, and the voice can be detected using a certain threshold value.

In the fifth invention, 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 the reference pattern. Further, as the recognition unit standard pattern here, a syllable, a phoneme, a word or the like as shown on page 5 of the document [1] can be used.

The recognition pattern generation unit generates a recognition pattern in which the reference pattern and the standard pattern are combined.

In the comparison and collation section, the feature vector time series and the recognition pattern are compared and collated by defining the end frame, and the standard pattern constituting the recognition pattern giving the maximum similarity is obtained as the recognition result. Here, as a method of comparison and collation with a word pattern, as shown on page 154-165 of the document [1], a feature vector time series is held as a standard pattern and the similarity is calculated based on the distance between feature vectors. A method of calculating and comparing and matching based on dynamic programming,
There is a method of comparison and matching based on the HMM as described in the document [4]. Further, as a method of comparison and collation with a reference pattern, continuous speech recognition by frame synchronous DP matching as shown in the literature [5],
H as shown on pages 40-50 of document [6]
A continuous speech recognition algorithm by MM can be used.

The reference similarity calculator calculates the similarity by comparing and collating the feature vector time series and the reference pattern with the end frame defined, and obtains the maximum similarity as the reference similarity. As a method of comparison and collation with a reference pattern, continuous speech recognition by frame synchronization DP matching as shown in the document [5] or HMM as shown on pages 40-50 of the document [6] is used. Continuous speech recognition algorithms can be used.

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

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

The influence of the speaker or the environment appearing in the input voice appears in both the case of obtaining the similarity with the standard pattern and the case of obtaining the reference similarity used for correction. In the corrected similarity thus obtained, the influences of the speaker and the environment are offset.

In the second invention, after calculating the similarity between the feature vector time series and a plurality of standard patterns registered in advance by setting the end frame and obtaining the start frame corresponding to the maximum similarity, Then, extract the part between the start frame and the end frame from the feature vector time series,
The reference similarity is calculated. Since the reference similarity is redone for each end frame, there is a problem that the amount of calculation increases. In the fifth aspect of the invention, the similarity by comparison and matching between the reference pattern and the recognition pattern obtained by combining the standard patterns is corrected by using the reference similarity by comparison and matching with the reference pattern. The result of the comparison and matching between the input voice and the recognition pattern is divided into the part corresponding to the standard pattern to be recognized and the part corresponding to the reference pattern. It means that word spotting is performed by associating the sub-sections with the standard pattern. Regarding the portion corresponding to the reference pattern, it is considered that substantially the same reference pattern is associated even in the calculation of the reference similarity, and the similarity has substantially the same value. Therefore, by correcting the similarity with the recognition pattern with the similarity with the reference pattern, the similarity with the standard pattern is corrected using the similarity with the reference pattern for the section corresponding to the standard pattern. This means that the corrected similarity is obtained. As a result, it is possible to perform word spotting in which the correction of the degree of similarity is applied by the comparison and collation with the fixed start point, the amount of calculation is small, and highly accurate word spotting can be performed.

[0039]

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

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

This speech recognition apparatus comprises 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 vector-to-vector similarity calculation unit 2 finds and outputs the vector-to-vector similarity D between the one-frame feature vector of the feature vector time series V and the feature vector forming the standard pattern registered in advance. is there.

The similarity correction unit 3 receives the inter-vector similarity D, finds the maximum value of the inter-vector similarity for one frame, and corrects the inter-vector similarity for each frame with the maximum value found. It is output as the corrected similarity F.

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

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

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

The voice detection unit 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. 1 will be described.

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

The feature vector time series V is input to the vector similarity calculation unit 2 and the similarity with the feature vectors forming a plurality of standard patterns registered in advance is calculated to obtain the vector similarity D. Become. As a method of obtaining the similarity between vectors, for example, 154- of Reference [1] is used.
A method based on inter-vector distance as described on page 161 can be used.

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

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

The corrected similarity C is input to the identification unit 5, the voice is identified using the corrected similarity, and the result is the recognition result A.
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 of the present invention, it is possible to obtain the corrected similarity not influenced by the speaker or the environment.

According to the third aspect, the reject unit 7 outputs the reject signal J when the corrected similarity C is smaller than a predetermined threshold value.

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

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

This voice recognition device comprises an analysis unit 11, a comparison and collation unit 12, a unit standard pattern storage unit 13, a reference similarity calculation unit 14, a similarity correction unit 15, and a voice detection unit 17.

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

The comparing and collating unit 12 determines the feature vector time series V
And the standard pattern registered in advance are compared and collated by defining the end frame E, the standard pattern giving the maximum similarity S is obtained as the recognition result A, and the starting frame T corresponding to the maximum similarity is obtained. is there.

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

The reference similarity calculation unit 14 combines the unit standard patterns registered in the unit standard pattern storage unit 13 in a predetermined order, from the feature vector time series V to the starting frame T. The section up to the end frame E is compared and collated with the feature vector time series, 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 the corrected similarity C.

The voice detection unit 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 converted into a feature vector time series V by feature analysis.

The feature vector time series V is compared and compared by the comparison and collation unit 12
Is input to a plurality of standard patterns registered in advance and the end frame E is defined and compared to calculate the similarity, and the standard pattern giving the maximum similarity S and the maximum similarity is obtained as the recognition result A. At the same time, the start frame T corresponding to the maximum similarity is obtained. Here, as a method of calculating the similarity, a method based on DP matching as shown in documents [1] and [5], or based on HMM as shown in documents [4] and [6] is used. Any method can be used.

The feature vector time series V is input to the reference similarity calculation unit 14 and is stored in the recognition unit standard pattern storage unit 13 and a portion extracted from the feature vector time series between the start frame T and the end frame E. A plurality of standard patterns obtained by combining the unit standard patterns that have been combined in a predetermined order are compared and collated, and the reference similarity R is output. As the recognition unit standard pattern here, a syllable as shown on page 5 of the 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 the correction method here, a method of obtaining a value based on the difference between S and R can be used.

In this way, according to the second invention, it is possible to obtain the corrected similarity not influenced by the speaker or the environment.

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

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

This voice recognition device 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 voice detection unit 27, a standard pattern storage unit 28, and recognition. A pattern generation unit 29 and a reference pattern generation unit 30 are provided.

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

The standard pattern storage unit 28 stores standard patterns.

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

The reference pattern generating section 30 is for generating a reference pattern Q in which the 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 uses the feature vector time series V.
And the recognition pattern P are compared and collated, and the standard pattern forming the recognition pattern giving the maximum similarity S is recognized as the recognition result A.
It is what you ask for.

The reference similarity calculator 24 determines the reference pattern Q.
And the feature vector time series V are compared and collated, 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 the corrected similarity C.

The voice detection section 27 recognizes the recognition result A when the corrected similarity C is larger than a predetermined threshold value.
Is detected.

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

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

The characteristic vector time series V is compared by the comparison and collation unit 22.
Is input to the recognition pattern P, the similarity is compared and collated, and the standard pattern forming the recognition pattern that gives the maximum similarity S and the maximum similarity is obtained as the recognition result A. Here, the similarity calculation method For example, a method based on DP matching as shown in documents [1] and [5] or H as shown in documents [4] and [6].
An MM-based method can be used.

Further, the feature vector time series V is input to the reference similarity calculation section 24 and compared and collated 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 value, the voice detection unit 28 detects the voice as a voice and outputs the recognition result A.

As described above, according to the fifth aspect of the invention, it is possible to obtain the corrected similarity not affected by the speaker or the environment and perform the word spotting.

[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 set to the maximum value of the vector-to-vector similarity of each frame or the input voice and the reference voice. By correcting 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 reject and word spotting can be realized.

[Brief description of drawings]

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

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

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

[Explanation of symbols]

 1 analysis unit 2 vector similarity calculation unit 3 similarity correction unit 4 similarity accumulation unit 5 discrimination unit 6 reject unit 7 voice detection unit 11 analysis unit 12 comparison 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 collation 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

Claims (5)

[Claims] 1. An analyzer for converting an input voice signal into a time series of frames of feature vectors, and a vector between a feature vector forming a standard pattern registered in advance and a feature vector of the input voice signal. An inter-vector similarity calculation unit that obtains a similarity for each frame of an input voice signal, and a frame-corrected similarity that is corrected using the maximum value of the inter-vector similarity in each frame of an input voice signal A voice recognition device comprising a similarity correction unit, a similarity accumulation unit that accumulates the frame correction similarities to obtain a correction similarity, and an identification unit that identifies a voice based on the corrected similarity. 2. An analysis unit for converting an input voice signal into a time series of feature vector frames, and the feature vector time series and a standard pattern registered in advance are compared and collated by defining an end frame, A comparison and collation unit that obtains a standard pattern giving a similarity as a recognition result and a starting frame corresponding to the maximum similarity, a unit standard pattern storage unit that holds a standard pattern of a recognition unit, and a standard pattern of the recognition unit Is obtained by the reference similarity calculation unit and the comparison and collation unit, which obtains the maximum value of the similarity between the standard pattern obtained by combining in a predetermined order and the feature vector time series from the start frame to the end frame as the reference similarity. And a similarity correction unit that obtains a corrected similarity by correcting the obtained similarity with the reference similarity. Recognition device. 3. The voice recognition device according to claim 1, further comprising a reject unit that outputs a reject signal when the corrected similarity obtained by the similarity correction unit is smaller than a predetermined threshold value. 4. The voice recognition device according to claim 1, further comprising a voice detection unit that detects a voice when the corrected similarity obtained by the similarity correction unit is larger than a predetermined threshold value. 5. An analysis unit for converting an input voice signal into a time series of feature vector frames, a standard pattern storage unit for holding a standard pattern to be recognized, and a unit standard pattern for holding a standard pattern of a recognition unit. A storage unit, a reference pattern generation unit that combines the standard patterns of the recognition units in a predetermined order to form a reference pattern, and a recognition pattern generation unit that combines the reference pattern and the standard pattern to be recognized as a recognition pattern And a comparison and collation unit that determines and compares the feature vector time series and the recognition pattern by defining an end frame, and obtains a standard pattern forming a recognition pattern that gives the maximum similarity as a recognition result, the reference pattern, and the reference pattern. Reference similarity that determines the maximum value of similarity when the end frame is defined as the end and compared and compared The correction similarity obtained by the calculation unit and the similarity correction unit and the similarity correction unit that obtains the corrected similarity obtained by correcting the maximum similarity obtained by the comparison and matching unit using the reference similarity is calculated in advance. A voice recognition device having a voice detection unit that detects a voice when the voice is larger than a predetermined threshold.