JP2003044078A - Voice recognizing device using uttering speed normalization analysis - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the recognition rate by appropriately extracting features corresponding to a uttering speed by controlling an analytic method corresponding to the uttering speed of an input voice. SOLUTION: Concerning voices inputted from voice input means 1 and 6, the uttering speeds of the input voices are outputted by uttering speed output means 2 and 7, an analytic control parameter is determined by analytic control parameter calculations 4 and 8 by using the uttering speed and according to that analytic parameter, the features of the input voices are extracted by feature extracting means 4 and 9.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to speech recognition,
In particular, it relates to voice recognition using vocalization speed normalization for voice analysis.

[0002]

2. Description of the Related Art Conventionally, a method used in speech recognition to reduce the influence of variations in utterance speed and improve the recognition rate will be described below by taking speech recognition using an HMM as an example.

First, when investigating the nature of speech, the speech waveform is analyzed to extract the features related to the frequency spectrum. Specifically, a short-time spectrum is obtained by applying a time window to the speech waveform and performing a discrete Fourier transform by using a fast Fourier transform algorithm. A voice section cut out with a time window is called a frame, and a cycle of moving this section is called a frame cycle (interval). Generally, the time window length is about 20 ms, and the frame period (interval) is constant.

The short-time spectrum of speech can be decomposed into a component (spectrum envelope) that gently changes with frequency and a component (spectrum fine structure) that finely changes with frequency. By examining this, the resonance / anti-resonance characteristics of the vocal tract and the sound source The features included in the voice such as the periodicity of can be captured. In general, the feature vector used for speech recognition has about 40 dimensions.

In HMM speech recognition, the time series of speech feature vectors is considered to change stochastically, and each word or phoneme is expressed by a stochastic state transition model (HMM) for recognition. FIG. 8 shows an example of HMM. It is a figure. As shown in FIG.
The MM is represented by each state and the transitions between them. Appearance probabilities and transition probabilities are assigned to each arc, and symbols corresponding to speech and temporal variations are probabilistically represented by these probabilities. In order to prepare an HMM with a high recognition rate in speech recognition, learning data in which a symbol sequence corresponding to a statistically sufficient amount of time series of speech feature vectors is used as a set and the symbol sequence is the highest. "Learning" is performed to estimate the probability parameter of the HMM so as to generate with probability.

At the time of recognition, HMMs are prepared by the number of words or phonemes to be recognized, and for all HMMs, the probability that the HMM outputs the symbol sequence of the input speech is calculated, and the HMM with the highest probability is calculated. The output word or phoneme is used as the recognition result.

Since the probability parameter of the HMM is estimated by using the learning data, the recognition performance is poor when the utterance speed of the input voice is far from the distribution of the utterance speed of the learning data. On the other hand, measuring the speaking speed of the input voice,
A method has been proposed in which recognition performance is improved by limiting the duration of a speech symbol sequence based on the speaking rate.

An example of a conventional voice recognition device is disclosed in Japanese Unexamined Patent Publication No.
No. 113298. The configuration is shown in FIG.
Shown in. In FIG. 9, the utterance speed detection unit 101 detects the utterance speed 103 from the peak of the voice power outline spectrum. The speech symbolization unit 102 analyzes the input speech and outputs a symbol sequence 104 symbolized by vector quantization. The speech recognition unit 105 based on the HMM method receives the speech rate 103 output by the speech rate detection unit 101 and the symbol sequence 104 output by the speech symbolization unit 102 as input, and limits the duration of the HMM state using the speech rate. While performing, the probability calculation of each word HMM in the word HMM database 106 and the symbol sequence 104 is performed, and the word with the highest probability is output as the recognition result 108.

[0009]

However, the above-mentioned conventional apparatus has a problem that it is not possible to appropriately extract the features in accordance with the speech production speed of the voice. The reason is that the conventional voice recognition device uses a constant frame interval or a frame interval that is an integral multiple thereof for voice analysis.

Therefore, a technical problem of the present invention is to provide a voice recognition device capable of performing an appropriate feature amount extraction in accordance with the utterance speed of an input voice.

Another technical object of the present invention is to provide a voice recognition device capable of reducing the processing time.

[0012]

According to the present invention, at least one of a learning voice and a recognition voice is analyzed by expanding and contracting an analysis frame interval based on a utterance speed. The device is obtained.

Further, according to the present invention, in the voice recognition device, the learning voice is analyzed by expanding and contracting an analysis frame interval based on a utterance speed, and the recognition voice is analyzed. A recognition device is obtained.

Further, according to the present invention, in the voice recognition device, the recognition voice is analyzed by expanding and contracting an analysis frame interval based on a utterance speed, and the learning voice is analyzed. A recognition device is obtained.

Further, according to the present invention, in the voice recognition device, the voice recognition device is characterized in that both the learning voice and the recognition voice are analyzed by expanding and contracting the analysis frame interval based on the utterance speed. can get.

Further, according to the present invention, in any one of the speech recognition devices, the calculation of an analysis frame interval used when analyzing the recognition voice by expanding and contracting the analysis frame interval based on the utterance speed, A voice recognition device is obtained which uses the utterance speeds of the learning voice and the recognition voice.

Further, according to the present invention, in any one of the voice recognition devices, the calculation of the analysis frame interval used when the recognition voice is analyzed by expanding and contracting the analysis frame interval based on the utterance speed. A voice recognition device is obtained which uses the utterance speed of only the recognition voice.

Further, according to the present invention, in any one of the voice recognition devices, the input voice is analyzed to extract a feature, and the extracted feature amount, a standard pattern prepared in advance and a time axis are associated with each other. It is possible to obtain a voice recognition device characterized in that the analysis frame interval is expanded or contracted based on the utterance speed calculated by performing the above to analyze the input voice.

Further, according to the present invention, in any one of the above speech recognition devices, a speaker utters a prepared sentence prepared in advance, and the utterance calculated from the number of syllables and the utterance duration included in the sentence. A voice recognition device is obtained which analyzes an input voice by expanding and contracting an analysis frame interval based on speed.

Further, according to the present invention, in the voice recognition device, when the recognition voice is analyzed by expanding and contracting the analysis frame interval based on the utterance speed, the analysis frame calculated from the utterance speed of the immediately preceding utterance. A voice recognition device is obtained which is characterized by using intervals.

Further, according to the present invention, in any one of the above speech recognition devices, the calculation of the analysis frame interval used when analyzing the input voice by expanding and contracting the analysis frame interval based on the utterance speed, A voice recognition device is obtained which is characterized by using an average speaking rate for each speaker.

Further, according to the present invention, in any one of the above speech recognition devices, the calculation of the analysis frame interval used when analyzing the input voice by expanding and contracting the analysis frame interval based on the utterance speed A voice recognition device is obtained which is characterized by using an average utterance speed for each utterance.

Further, according to the present invention, in any one of the above speech recognition devices, vowels are used for calculating an analysis frame interval used when analyzing an input voice by expanding and contracting the analysis frame interval based on the utterance speed. A voice recognition device is obtained which uses the vowel production speed of vowels and the consonant production speed of consonants.

Further, according to the present invention, in any one of the voice recognition devices, the analysis frame interval used when analyzing the input voice by expanding and contracting the analysis frame interval based on the utterance speed is different from that of the stationary part. A voice recognition device is obtained which is characterized by being calculated separately in the stationary unit.

According to the present invention, in any one of the voice recognition devices, whether or not to analyze the input voice by expanding and contracting the analysis frame interval based on the utterance speed is determined by using a threshold for the utterance speed. A voice recognition device characterized by controlling is obtained.

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION Next, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to the first embodiment of the present invention. Referring to FIG. 1, a voice recognition device 10 according to a first exemplary embodiment of the present invention includes a learning voice input unit 1 for inputting a voice used for acoustic model learning, a utterance speed output unit 2 for a learning voice, and a utterance. Analysis control parameter calculation (functional section) 3 for calculating analysis frame intervals based on speed, feature extraction means 4
A recognition voice input means 6, a utterance speed output means 7 for the recognized voice, an analysis control parameter calculation (functional unit) 8 for calculating an analysis frame interval based on the utterance speed, a feature extraction means 9, and an acoustic model learning means. 5, a recognition unit 11, and a recognition result output unit 12 are included.

The analysis control parameter calculations 3 and 8 receive the outputs of the speech production speed output means 2 and 7, respectively, and the analysis control parameters for the feature extraction means 4 and 9 to extract an appropriate feature amount corresponding to the production speed. As, the analysis frame interval is calculated based on the speaking rate. The feature extracting means 4 and 9 receive the analysis control parameter and the input voice as input, analyze the input voice according to the analysis control parameter, and perform feature extraction.

The acoustic model learning means 5 is the feature extraction means 4
The parameters of the acoustic model are estimated by using the feature amount of the learning voice output in step S3, and the pattern referred to in the acoustic process of the voice recognition process is created.

The recognizing means 11 receives the acoustic model created by the acoustic model learning means 5 and the feature amount of the recognition voice output from the feature extracting means 9 as input and performs matching to output the most probable correct answer candidate to the recognition result output means. At 12, the result is output as a recognition result.

The operation and effect of the first embodiment of the present invention will be described.

In the present embodiment, the utterance speed of the input voice is used for the analysis control parameter calculation, and the analysis is performed in accordance with the analysis control parameter, whereby it is possible to perform the appropriate feature extraction according to the utterance speed. The recognition rate is improved by inputting the acoustic model learned using the thus obtained learning voice feature amount and the recognition voice feature amount to the recognition means.

Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 2 is a block diagram showing the structure of a speech recognition apparatus according to the second embodiment of the present invention.

Referring to FIG. 2, a speech recognition apparatus 20 according to the second embodiment of the present invention outputs a learning speech utterance speed to an input of an analysis control parameter calculation section 8 constituting a recognition speech processing section. The point that the output of the means 2 is also used is different from the recognition voice processing unit in the first embodiment shown in FIG.

Next, the function and effect of the second embodiment of the present invention will be described.

In the second embodiment of the present invention, in addition to the effect of the first embodiment, not only the recognition voice information but also the learning voice information is used for feature extraction of the recognition voice. , It is possible to perform voice analysis on the recognition voice in accordance with the environment in which the acoustic model is created.

Next, a third embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 3 is a block diagram showing the structure of a speech recognition apparatus according to the third embodiment of the present invention.

Referring to FIG. 3, the speech recognition apparatus 30 according to the third embodiment of the present invention includes a learning speech processing unit,
The difference is that the utterance speed output means 2 and the analysis control parameter calculation 3 which constitute the learning voice processing unit in the first embodiment shown in FIG. 1 do not exist. As a result, the learning voice feature extraction unit 4 performs the conventional analysis.

Next, the effect of the third embodiment of the present invention will be described.

In the third embodiment of the present invention, when compared with the first embodiment, the processing of the learning voice processing unit is the same as the conventional one, and it is not necessary to measure the utterance speed and calculate the analysis parameter. The processing time can be reduced.

Next, a fourth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 4 is a block diagram showing the structure of a speech recognition apparatus according to the fourth embodiment of the present invention.

Referring to FIG. 4, a speech recognition apparatus 40 according to the fourth embodiment of the present invention has a learning speech processing unit,
The difference is that the analysis control parameter calculation 3 forming the learning speech processing unit in the second embodiment shown in FIG. 2 does not exist. As a result, the learning voice feature extraction unit 4 performs the conventional analysis.

Next, the function and effect of the fourth embodiment of the present invention will be described.

In the fourth embodiment of the present invention, when compared with the second embodiment, the learning voice processing unit does not need to calculate the analysis parameter and can reduce the processing time.

Next, a fifth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 5 is a block diagram showing the structure of a speech recognition apparatus according to the fifth embodiment of the present invention.

Referring to FIG. 5, a speech recognition apparatus 50 according to the fifth embodiment of the present invention has a recognition speech processing section as the recognition speech processing section in the first embodiment shown in FIG. The difference is that the utterance speed output means 7 and the analysis control parameter calculation 8 do not exist. As a result, the recognition voice feature extraction means 9 performs the conventional analysis.

Next, the function and effect of the fifth embodiment of the present invention will be described.

In the fifth embodiment of the present invention, when compared with the first embodiment, the processing of the recognition voice processing unit composed of the recognition voice input means 6 and the feature extraction means 9 is the same as the conventional one. The processing time can be reduced without the need for measurement of vocalization speed and calculation of analysis parameters.

Next, a sixth embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 6 is a block diagram showing the structure of a speech recognition apparatus according to the sixth embodiment of the present invention.

Referring to FIG. 6, in a voice recognition device 60 according to the sixth embodiment of the present invention, a recognition voice processing unit is
In addition to the configuration of the recognition voice processing unit in the first embodiment shown in FIG. 1, it is different in that it has a parameter storage / readout unit 13.

The parameter storing / reading means 13 reads the analysis control parameter calculated and held using the immediately preceding utterance, and the analysis control parameter calculated using the current utterance for the feature extraction of the next utterance. Memorize.

The feature quantity extraction means 9 receives as input the analysis control parameters read by the parameter storage / readout means 13 and calculated using the immediately preceding utterance and the current recognition voice input to the recognition voice input means 6. , The feature extraction of the recognition voice is performed according to the analysis control parameter.

Next, the function and effect of the sixth embodiment of the present invention will be described.

In the sixth embodiment of the present invention, the feature of the current recognition voice is extracted by using the analysis control parameter calculated by using the immediately preceding utterance, whereby the utterance speed is measured and the analysis control parameter is calculated. The processing time required can be reduced.

FIG. 7 shows a specific example of the present invention.
This will be described with reference to (a) and FIG. 7 (b). This specific example corresponds to the second embodiment of the present invention.

According to this method, it is possible to perform appropriate feature extraction according to the utterance speed while measuring the utterance speed and freely changing the analysis control parameter for the learning voice and the recognition voice. 7 (a), when the preliminary feature extraction means 21 and the vocalization speed calculation means 22 are used as the vocalization speed output means, or as shown in FIG. 7 (b), the prescribed sentence vocalization time length measurement (functional section ) 25 and the speaking speed calculation means 26 are used as an example.

First, the method for extracting the characteristics of the learning voice will be described. The learning voice input from the microphone by the learning voice input unit 1 is input to the speaking speed output unit 2 and the feature extraction unit 4.

Here, as an example, the case where the utterance speed output means 2 is composed of the preliminary feature extraction means 21 and the utterance speed calculation means 22 as shown in FIG. 7A will be described.

The preliminary feature extracting means 21 performs feature extraction for measuring the speaking speed of the input voice. The preliminary feature extraction means 21 performs a conventional analysis on the input voice to perform feature extraction. The utterance speed calculation means 22 associates the feature amount extracted by the preliminary feature extraction means 21 with the HMM of each phoneme prepared in advance on the time axis, and calculates and outputs the utterance speed of the input speech. Examples of units for calculating the vocalization rate include one vocalization, a vowel part, a consonant part, a stationary part, and a non-stationary part.

Next, as another example, as shown in FIG. 7B, the utterance speed output means 2 is composed of a prescribed sentence utterance time length measurement (function part) 25 and a utterance speed calculation means 26. explain. In the prescribed sentence utterance time length measurement 25, the voice that utters the prepared prescribed sentence is input, and the time required for utterance is measured. Speech rate calculation means 2
In 6, the utterance dissimilarity of the input voice is calculated and output from the number of phonemes contained in the prescribed sentence and the prescribed sentence utterance duration measured by the prescribed sentence utterance duration measurement 25.

Referring again to FIG. 2, the utterance speed output from the utterance speed output means 2 is input to the analysis control parameter calculation 3, and the feature extraction means 4 obtains an appropriate feature amount according to the utterance speed of the input voice. As an analysis control parameter for extraction, an analysis frame interval based on the speaking rate is calculated.
A calculation example of the analysis frame interval is shown in the following mathematical formula 1.

[0067]

[Equation 1] Here, “I” is an analysis frame interval used in the feature extraction means 4, Io is a constant analysis frame interval used in normal analysis, and C is calculated from the utterance speed output by the utterance speed output means 2. A coefficient, for example, the average utterance rate of all utterances, R is a unit for normalization, for example, one utterance, a vowel part,
It is the vocalization speed of the consonant part, the stationary part, and the non-constant part. s is
Acts as a switch to switch the analysis method. s is 0
If so, I becomes equal to Io and becomes equal to the conventional analysis method, and if s is 1, "I" becomes the analysis frame interval normalized by the vocalization rate, and the vocalization normalization analysis is performed. . In order not to be the same as the conventional analysis method, it is assumed that s is not 0 throughout all vocalizations. By calculating this equation 1 according to the unit to be normalized, the vowel production rate for vowels, the consonant production rate for consonants, the stationary section production rate for stationary parts, For the non-stationary part, the analysis frame interval using the vocalization rate of the non-stationary part can be obtained. For example, the value of s may be set to 1 throughout all utterances, and the utterance rate normalization analysis may be performed on all utterances. If it is 1, the vocalization speed normalization analysis is performed, and if it is slow, it is set to 0, and the conventional analysis may be performed.

The feature extraction means 4 receives the analysis control parameters output by the analysis parameter calculation (function part) 3 and the learning voice input by the learning voice input means 1, and outputs the learning voice according to the prayer control parameters. Feature extraction is performed.

Next, a method for extracting the features of the recognized organ voice will be described. The recognition voice input means 6 is the same as the learning voice input means 1, and the utterance speed output means 7 is the same as the utterance speed output means 2.
The feature extracting means 9 performs the same operation as the feature extracting means 4.
In the analysis control parameter calculation (functional unit) 8, a coefficient calculated from the utterance speed of the learning voice output by the utterance speed output means 2 to C in the above formula 1, for example, the average utterance speed of all utterances, utterance to R The utterance speed of the recognition voice output by the speed output means 7, for example, the utterance speed of one utterance, a vowel part, a consonant part, a stationary part, and a non-stationary part is used.

The feature extraction means 9 receives the analysis control parameters output from the analysis parameter calculation (function part) 8 and the recognition voice input from the recognition voice input means 6 as input, and the features of the recognition voice according to the analysis control parameters. Extract.

The acoustic model learning means 5 receives the characteristic amount output from the characteristic extracting means 4 as an input, estimates the parameters of the acoustic model, and creates a pattern to be referred to in the acoustic processing of the speech recognition processing.

The recognizing means 11 receives the acoustic model created by the acoustic model learning means 5 and the feature amount of the recognition voice output from the feature extracting means 9 as input and performs matching, and outputs the most probable correct answer candidate as a recognition result. The means 12 outputs it as a recognition result.

[0073]

As described above, according to the present invention,
By using the analysis control parameters calculated from the vocalization speed of the input speech, the vocalization speed normalized analysis is performed in order to extract the appropriate features according to the vocalization speed. It is possible to provide a voice recognition device with improved performance.

Further, according to the present invention, since the analysis control parameter calculated by using the immediately preceding utterance is held / read out to extract the feature of the current recognition voice, the immediately preceding utterance can be obtained with respect to the current utterance. Since the feature extraction can be sequentially performed using the utterance information, it is possible to provide a speech recognition device that can reduce the processing time.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a voice recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a voice recognition device according to a second embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration of a voice recognition device according to a third embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 5 is a block diagram showing a configuration of a voice recognition device according to a fifth embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a voice recognition device according to a sixth embodiment of the present invention.

7 (a) and 7 (b) are block diagrams respectively showing a concrete configuration example of a speech production speed output means in the second exemplary embodiment of the present invention.

FIG. 8 is a diagram showing an example of a conventional HMM.

FIG. 9 is a block diagram showing a configuration of a voice recognition device in a conventional example.

[Explanation of symbols]

1 Voice input means for learning 2 Speech rate output means 3 Analysis control parameter calculation 4 Feature extraction means 5 Acoustic model learning means 6 Speech input means for recognition 7 Speech rate output means 8 Analysis control parameter calculation 9 Feature extraction means 10, 20, 30, 40, 50, 60 Speech recognition device 21 preliminary feature extraction means 22 Speech rate calculation means 25 Regular sentence vocalization time measurement 26 Speech rate calculation means 100 input terminals 101 Speech rate detector 102 Speech symbolization section 103 Speech rate 104 symbol series 105 HMM based speech recognition unit 106-word HMM database 107 HMM parameters 108 recognition result

Claims (14)

[Claims] 1. A voice recognition device, wherein at least one of a learning voice and a recognition voice is analyzed by expanding and contracting an analysis frame interval based on a utterance speed. 2. The voice recognition device according to claim 1, wherein
A voice recognition apparatus, characterized in that the learning voice is analyzed by expanding and contracting an analysis frame interval based on a utterance speed, and the recognition voice is analyzed. 3. The voice recognition device according to claim 1, wherein
A speech recognition apparatus, characterized in that the recognition speech is analyzed by expanding and contracting an analysis frame interval based on a utterance speed to analyze the learning speech. 4. The voice recognition device according to claim 1,
A voice recognition device, wherein both the learning voice and the recognition voice are analyzed by expanding / contracting analysis frame intervals based on the utterance speed. 5. The speech recognition apparatus according to claim 3, wherein the learning voice is used to calculate an analysis frame interval used when analyzing the recognition voice by expanding and contracting the analysis frame interval based on the utterance speed. And a voice recognition speed of the recognition voice is used. 6. The speech recognition apparatus according to claim 3, wherein the recognition voice is used for calculating an analysis frame interval used when analyzing the recognition voice by expanding and contracting the analysis frame interval based on a utterance speed. A voice recognition device characterized by using a speaking speed of only voice. 7. The voice recognition apparatus according to claim 1, wherein the input voice is analyzed to perform feature extraction, and the extracted feature amount and a standard pattern prepared in advance. A voice recognition device characterized in that an input voice is analyzed by expanding and contracting an analysis frame interval based on a utterance speed calculated by performing time axis association. 8. The voice recognition apparatus according to claim 1, wherein the speaker utters a prepared sentence prepared in advance, and the number of syllables and utterance duration included in the sentence. A speech recognition apparatus, which analyzes an input speech by expanding and contracting an analysis frame interval based on the utterance speed calculated from. 9. The voice recognition device according to claim 3,
A speech recognition apparatus characterized by using an analysis frame interval calculated from a utterance speed of an immediately preceding utterance when analyzing the recognition voice by expanding and contracting an analysis frame interval based on a utterance speed. 10. The speech recognition apparatus according to claim 1, wherein an analysis frame interval used when analyzing the input voice by expanding or contracting the analysis frame interval based on a speaking rate. A voice recognition device characterized by using an average speaking speed for each speaker for calculation. 11. The voice recognition device according to claim 1, wherein an analysis frame interval used when analyzing the input voice by expanding or contracting the analysis frame interval based on a speaking rate. A voice recognition device characterized by using an average utterance speed for each utterance for calculation. 12. The speech recognition apparatus according to claim 1, wherein an analysis frame interval used when analyzing the input voice by expanding or contracting the analysis frame interval based on the utterance speed. A voice recognition device characterized in that a vowel uses a vowel vocalization speed and a consonant uses a consonant vocalization speed for calculation. 13. The voice recognition device according to claim 1, wherein an analysis frame interval used when expanding and analyzing the analysis frame interval based on a utterance speed of the input voice is used. A voice recognition device characterized by being calculated separately for a stationary part and a non-stationary part. 14. The voice recognition apparatus according to claim 1, wherein whether or not to analyze the input voice by expanding or contracting an analysis frame interval based on the voice production speed is determined with respect to the voice production speed. A voice recognition device characterized by being controlled using a threshold value.