JP2003345383A - Method, device, and program for voice recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device which reduces the calculation volume of similarities in the case of matching processing to quickly and accurately recognize voice. <P>SOLUTION: A voice recognition device 100 is provided with a voice analysis part 105 for extracting the feature quantity of a voice signal with respect to each frame, a key word model database 106 wherein key word HMMs representative of feature quantity patterns of a plurality of key words to be recognized are stored, a similarity calculation part 107 for calculating similarity of the feature quantities of respective frames on the basis of the feature quantity of each extracted frame and key word HMMs, a matching processing part 108 for performing matching processing on the basis of the calculated similarity between each frame and each key word HMM and similarities of unnecessary words which don't constitute preliminarily set key words, and a discrimination part 109 for discriminating a key word included in a spoken voice on the basis of the matching processing. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to an HMM (Hidden M).
arkov Models) method, and more specifically, it belongs to the technical field of recognizing keywords from spoken speech.

[0002]

2. Description of the Related Art Currently, a voice recognition device for recognizing a voice uttered by a human being has been developed. In such a voice recognizing device, when a human utters a voice of a predetermined phrase, the voice of the phrase is input from the input signal. To recognize.

Further, when such a voice recognition device is applied to various devices such as a vehicle-mounted navigation device and a personal computer, the device inputs various information without requiring a manual operation such as a keyboard or switch selection. Will be able to.

Therefore, even in a working environment where a person uses both hands while using a navigation device while driving a car, the operator can input desired information into the device. ing.

HM is a typical example of such speech recognition.
A method for performing speech recognition using a probabilistic model called M (Hidden Markov Model) (hereinafter, simply referred to as speech recognition)
There is.

In this voice recognition, the pattern of the feature quantity of the uttered voice is matched with the pattern of the feature quantity of the voice showing the words and phrases of the recognition candidates (hereinafter referred to as recognition target words (keywords)) which are the prepared keywords. As a result, voice recognition is performed.

Specifically, in this voice recognition, a speech amount (input signal) input at predetermined time intervals is analyzed to extract a feature amount, and the feature amount of this input signal is stored in a database in advance. Data of the feature amount of the recognition target word indicated by the stored HMM and the matching rate, that is,
The probability that the feature quantity of the input signal is feature quantity data (hereinafter referred to as similarity degree) is calculated, and the similarity degrees of all utterance voices are integrated, and the recognition target having the highest integrated similarity degree is calculated. The word is fixed as a recognition result.

As a result, in this voice recognition, it is possible to perform voice recognition of a predetermined word / phrase from an input signal which is a spoken voice.

The HMM is a statistical signal source model represented as a group of transition states, and indicates a feature amount of speech such as a keyword to be recognized in advance. Also, this H
The MM is adapted to collect a plurality of voice data in advance and generate based on these voice data.

In such speech recognition, it is important how to extract a keyword portion which is a recognition target word included in a speech voice.

Generally, in addition to keywords, the utterance voice includes
It often contains unnecessary words (words such as "er" and "da" that are added before and after the recognition target word) that are unnecessary words when recognizing in advance. In principle, it is formed by unnecessary words and keywords sandwiched between the unnecessary words.

Conventionally, in the case of performing voice recognition, a method called word spotting (hereinafter simply referred to as word spotting voice recognition) which is performed by recognizing a keyword as a target of voice recognition is generally used.

In this word spotting voice recognition, the utterance voice to be recognized is HM indicating a keyword model.
In addition to M, an HMM indicating a model of an unnecessary word (hereinafter referred to as a garbage model) is prepared, and speech recognition is performed by recognizing a keyword model, a garbage model, or a combination thereof having the highest degree of similarity in feature quantity. It has become.

That is, this word spotting voice recognition recognizes a keyword model, a garbage model, or a combination thereof having the highest similarity in feature amount based on the accumulated similarity, and the utterance voice includes the keyword. In that case, the keyword is output as a recognition result.

As a method of constructing the unnecessary word model in the case of performing speech recognition as described above, there is a method of using a probabilistic model called a filler model (hereinafter simply referred to as a filler model).

As shown in FIG. 5, the filler model is a model in which the connection relationships of all vowels and consonants that can be connected to model all voices are expressed by a network. In order to realize spotting, it is necessary to connect filler models before and after the keyword model.

In other words, in the filler model, all recognizable patterns, specifically, the matching of the feature amount of the uttered speech to be recognized and the feature amount of each phoneme is calculated to calculate the phoneme connection of the uttered voice. However, it is not necessary to recognize it based on the cumulative similarity (cumulative likelihood) obtained by accumulating the similarity of each subword that is a unit that constitutes a word such as phoneme, phoneme, or syllable in all connection patterns considered as a language. It is designed to recognize words.

[0018]

However, in such a voice recognition device, since the unnecessary word is recognized, matching between the characteristic amount of the uttered voice and the data of each characteristic amount such as a phoneme that can be a constituent element of the unnecessary word. Since the processing is performed, the amount of calculation becomes enormous, and there is a problem in that the load of calculation processing is applied.

Specifically, the matching process is performed by calculating the similarity (likelihood) indicating the similarity between the feature amount of the uttered voice and the feature amount data that can be a constituent element of the unnecessary word, and calculating the maximum similarity. Since it is determined that the word should be recognized as an unnecessary word having feature amount data that increases, in the case of the Japanese filler model, the feature amount of the uttered speech is It is necessary to calculate the degree of similarity with the feature amount data of all syllables.

Therefore, the above-mentioned speech recognition apparatus has a problem in that the calculation amount becomes enormous due to the calculation of the degree of similarity for each feature amount data, so that the load of calculation processing is increased.

In particular, the above speech recognition apparatus assumes a language model in which filler models for recognizing unnecessary words are connected before and after a keyword to be recognized. Since a matching process between the amount and each feature amount data is performed, a larger amount of calculation is required.

The present invention has been made in view of the above problems, and its object is to reduce the calculation amount of the similarity when performing the matching process, and to perform the voice recognition accurately and at high speed. To provide a recognition device.

[0023]

In order to solve the above-mentioned problems, the invention according to claim 1 is a voice recognition device for recognizing a keyword included in an uttered voice, wherein the uttered voice is analyzed. Extraction means for extracting the uttered voice feature amount, which is the feature amount of the voice component of the uttered voice, and first storage means for storing in advance keyword feature amount data indicating the feature amount of the voice component of one or more of the keywords. And a probability that the uttered voice feature amount is the keyword based on the uttered voice feature amount extracted from at least a part of the utterance voice and the keyword feature amount data stored in the storage means. The calculation means for calculating each keyword probability shown and a preset value are used as unnecessary words in which at least a part of the speech section of the uttered speech does not form the keyword. Second storage means for storing as an unnecessary word probability indicating a probability of occurrence, and determining means for determining the keyword to be recognized included in the uttered voice based on the calculated keyword probability and the unnecessary word probability. It has a different configuration.

With this configuration, in the invention according to claim 1, the keyword probability indicating that the uttered voice feature amount is the keyword indicated by each keyword feature amount data is calculated, and the calculated keyword probability and the preset keyword probability are set. Based on the unnecessary word probability, the keyword to be recognized included in the spoken voice is determined.

Therefore, the unnecessary word and the keyword can be identified and the keyword to be recognized can be determined without calculating the characteristics of the characteristic amounts of the uttered voice characteristic amount and the unnecessary word characteristic amount data. The processing load when calculating the degree of similarity can be reduced, and the keywords included in the uttered voice can be easily and quickly recognized.

The invention described in claim 2 is the same as claim 1
In the voice recognition device according to the item (1), the extraction unit analyzes the uttered voice for each preset unit time to extract the uttered voice feature amount information, and the unnecessary word probability stored in the storage unit in advance. Is a case where the unnecessary word probability in the unit time is shown, the calculating means calculates each of the keyword probabilities based on the extracted speech feature amount for each unit time, and the determining means , The keyword to be recognized included in the uttered voice is determined based on the calculated keyword probability and the unnecessary word probability in the unit time.

With this configuration, in the invention according to claim 2, a keyword probability indicating that the uttered voice feature amount is a keyword indicated by the keyword feature amount data is calculated for each unit time, and the calculated keyword probability is calculated. And the keywords to be recognized included in the spoken voice are determined based on the unnecessary word probability in a preset unit time.

Therefore, the keyword to be recognized can be determined by calculating the degree of similarity for each linguistic sound expressed from the uttered speech in units of phonemes, or in units of subwords, which are units of words such as speech units. In addition, the unnecessary words and the keywords can be identified and the keywords to be recognized can be determined without calculating the characteristics of the characteristic features of the uttered voice characteristic amount and the unnecessary word characteristic amount data. The processing load at the time of calculation can be reduced, the keyword included in the uttered voice can be easily and quickly recognized, and erroneous recognition can be prevented.

The invention described in claim 3 is the same as that of claim 2.
In the voice recognition device according to the above item 3, the determination unit is indicated by each of the keyword feature amount data stored in the first storage unit based on the calculated keyword probability and the unnecessary word probability in the unit time. It is configured to calculate a combination probability indicating a probability of each combination of each keyword and the unnecessary word and determine the keyword to be recognized included in the uttered voice based on the combination probability.

With this configuration, in the invention according to claim 3, based on each keyword probability and unnecessary word probability calculated for each unit time based on each combination of the keyword and the unnecessary word indicated by the keyword feature amount data. Then, each combination probability of each keyword indicated by the keyword feature amount data and the unnecessary word is calculated, and the keyword to be recognized included in the uttered voice is determined based on the combination probability.

Therefore, the keyword included in the uttered voice can be determined while considering each combination of the unnecessary word similarity and each calculated keyword similarity. Therefore, the keyword included in the uttered voice can be easily and quickly determined. Not only can it be recognized, but erroneous recognition can be prevented.

According to a fourth aspect of the present invention, there is provided a voice recognition method for recognizing a keyword included in an uttered voice, which is a feature amount of a voice component of the uttered voice by analyzing the uttered voice. An extraction processing step of extracting a voice feature quantity, a first acquisition processing step of previously acquiring keyword feature quantity data indicating a feature quantity of a voice component of one or more of the keywords, and a voice of at least a part of the uttered voice. A calculation processing step of calculating each keyword probability indicating a probability that the utterance voice feature amount is the keyword, based on the utterance voice feature amount extracted from the section and the keyword feature amount data stored in the storage means; , A preset value is an unnecessary word probability indicating a probability that at least a part of the speech section of the uttered speech is an unnecessary word that does not form the keyword. A second acquisition processing step of acquiring Te,
And a determination processing step of determining the keyword to be recognized included in the uttered voice based on the calculated keyword probability and the unnecessary word probability.

With this configuration, in the invention according to claim 4, a keyword probability indicating that the uttered voice feature amount is a keyword indicated by each keyword feature amount data is calculated, and the calculated keyword probability and the preset keyword probability are set. Based on the unnecessary word probability, the keyword to be recognized included in the spoken voice is determined.

Therefore, the unnecessary word and the keyword can be identified and the keyword to be recognized can be determined without calculating the characteristics of the characteristic amounts of the uttered voice characteristic amount and the unnecessary word characteristic amount data. The processing load when calculating the degree of similarity can be reduced, and the keywords included in the uttered voice can be easily and quickly recognized.

The invention described in claim 5 is the same as claim 4
In the voice recognition method described in the above, in the extraction processing step, the uttered voice is analyzed at every preset unit time to extract the uttered voice feature amount information, and in the calculation processing step, the extracted The keyword probabilities are calculated based on the uttered speech feature amount for each unit time, and in the second acquisition processing step, an unnecessary word probability indicating an unnecessary word probability in the unit time is acquired in advance, and the determination processing is performed. In the step, the keyword to be recognized included in the speech is determined based on the calculated keyword probability and the unnecessary word probability in the unit time.

With this configuration, in the invention according to claim 5, the keyword probability indicating that the uttered voice feature amount is the keyword indicated by the keyword feature amount data is calculated for each unit time, and the calculated keyword probability is calculated. And the keywords to be recognized included in the spoken voice are determined based on the unnecessary word probability in a preset unit time.

Therefore, the keyword to be recognized can be determined by calculating the degree of similarity for each linguistic sound expressed in a phoneme unit or a subword unit which is a unit of a word such as a voice unit from the uttered speech. In addition, the unnecessary words and the keywords can be identified and the keywords to be recognized can be determined without calculating the characteristics of the characteristic features of the uttered voice characteristic amount and the unnecessary word characteristic amount data. The processing load at the time of calculation can be reduced, the keyword included in the uttered voice can be easily and quickly recognized, and erroneous recognition can be prevented.

The invention according to claim 6 is the same as claim 5
In the voice recognition method according to, in the determination processing step, based on the calculated keyword probability and the unnecessary word probability in the unit time, each keyword indicated by the acquired keyword feature amount data and the A combination probability indicating a probability of each combination with an unnecessary word is calculated, and the keyword to be recognized included in the uttered speech is determined based on the combination probability.

With this configuration, in the invention according to claim 6, the keyword similarity and the unnecessary word similarity calculated for each unit time based on each combination of the keyword and the unnecessary word indicated by the keyword feature amount data. Are accumulated to calculate a cumulative similarity, and the keyword to be recognized included in the uttered voice is determined based on the cumulative similarity.

Therefore, it is possible to determine the keyword included in the uttered voice while considering each combination of the unnecessary word similarity and each calculated keyword similarity, so that the keyword included in the uttered voice can be easily and quickly determined. Not only can it be recognized, but erroneous recognition can be prevented. According to a seventh aspect of the present invention, there is provided a voice recognition program for performing a voice recognition process of recognizing a keyword included in a uttered voice by a computer, wherein the computer analyzes the uttered voice to produce the uttered voice. Extracting means for extracting a uttered voice feature amount, which is a feature amount of the voice component, and a first obtaining means for obtaining in advance the keyword feature amount data indicating the feature amount of the voice component of one or more of the keywords; Calculate each keyword probability indicating the probability that the utterance voice feature amount is the keyword, based on the extracted utterance voice feature amount of at least a part of the voice section and the keyword feature amount data stored in the storage means. Calculating means for setting a preset value such that at least a part of the speech section of the uttered speech does not constitute the keyword. As second acquisition means for acquiring as an unnecessary word probability indicating a probability of being an unnecessary word, determination means for determining the keyword to be recognized included in the uttered voice based on the calculated keyword probability and the unnecessary word probability. It has a functioning function.

With this configuration, in the invention according to claim 7, based on each keyword probability and unnecessary word probability calculated for each unit time based on each combination of the keyword and the unnecessary word indicated by the keyword feature amount data. Then, each combination probability of each keyword indicated by the keyword feature amount data and the unnecessary word is calculated, and the keyword to be recognized included in the uttered voice is determined based on the combination probability.

Therefore, the unnecessary word and the keyword can be discriminated and the keyword to be recognized can be determined without calculating the characteristics of the characteristic amounts of the uttered voice characteristic amount and the unnecessary word characteristic amount data. The processing load when calculating the degree of similarity can be reduced, and the keywords included in the uttered voice can be easily and quickly recognized.

The invention according to claim 8 is the same as claim 7
In the voice recognition program according to claim 1, the computer, the extraction means for analyzing the uttered voice for each preset unit time to extract the uttered voice feature amount information, the uttered voice extracted for each unit time Calculation means for calculating each keyword probability based on a feature amount, second acquisition means for previously acquiring the unnecessary word probability in the unit time,
It is configured to function as a determining unit that determines the keyword to be recognized included in the spoken voice based on the calculated keyword probability and the unnecessary word probability in the unit time.

With this configuration, in the invention according to claim 8, a keyword probability indicating that the uttered voice feature amount is a keyword indicated by the keyword feature amount data is calculated for each unit time, and the calculated keyword probability is calculated. And the keywords to be recognized included in the spoken voice are determined based on the unnecessary word probability in a preset unit time.

Therefore, the keyword to be recognized can be determined by calculating the degree of similarity for each language sound expressed from the uttered speech in units of phonemes, or in units of subwords, which are units of words such as units of speech. In addition, the unnecessary words and the keywords can be identified and the keywords to be recognized can be determined without calculating the characteristics of the characteristic features of the uttered voice characteristic amount and the unnecessary word characteristic amount data. The processing load at the time of calculation can be reduced, the keyword included in the uttered voice can be easily and quickly recognized, and erroneous recognition can be prevented.

The invention described in claim 9 is the same as claim 8
In the speech recognition program according to claim 1, the computer, based on the calculated keyword probability and the unnecessary word probability in the unit time, between each keyword and the unnecessary word indicated by the keyword feature amount data acquired in advance The configuration is such that a combination probability indicating a probability of each combination is calculated, and based on the combination probability, it functions as a determination unit that determines the keyword to be recognized included in the uttered voice.

With this configuration, in the invention described in claim 9, based on each keyword probability and unnecessary word probability calculated for each unit time based on each combination of the keyword and the unnecessary word indicated by the keyword feature amount data. Then, each combination probability of each keyword indicated by the keyword feature amount data and the unnecessary word is calculated, and the keyword to be recognized included in the uttered voice is determined based on the combination probability.

Therefore, the keyword included in the uttered voice can be determined while considering each combination of the unnecessary word similarity and each calculated keyword similarity, so that the keyword included in the uttered voice can be easily and quickly determined. Not only can it be recognized, but erroneous recognition can be prevented.

[0049]

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

The embodiment described below is an embodiment in which the voice recognition device according to the present invention is applied.

First, a spoken language model using the HMM of this embodiment will be described with reference to FIG.

FIG. 1 is a diagram showing a spoken language model showing a recognition network using the HMM of this embodiment.

In this embodiment, the HMM as shown in FIG.
It is assumed that the model showing the recognition network using (hereinafter referred to as a spoken language model), that is, the spoken language model 10 including the keyword to be recognized.

The spoken language model 10 has a structure in which a model (hereinafter, referred to as an unnecessary word constituent element model) 12a and 12b called a garbage model, which is a unit for forming an unnecessary word, is connected before and after the keyword model 11. Keywords included in the uttered voice are displayed in the keyword model 11,
By matching the unnecessary words with the unnecessary word constituent models 12a and 12b, the keywords and the unnecessary words are identified, and the keywords included in the uttered speech are recognized.

In practice, in the present embodiment, the keyword model 11 represents a group of states that transit for each arbitrary section of the uttered speech, and a statistical signal source model in which a non-stationary signal source is represented by a connection of stationary signals. And the unnecessary word component models 12a and 12b are represented by fixed values set in advance in order to reduce the calculation amount as described later.

The HMM of the keyword model 11 (hereinafter referred to as a keyword HMM) is a state transition probability indicating a probability of transition of a state from a certain state to a vector (feature amount for each frame) observed when the state transits. It has two parameters of output probability that outputs the probability of (vector), and indicates the feature amount pattern of each keyword.

In general, a speech voice has acoustic fluctuations caused by various causes even if the same word or syllable is present. Therefore, if the speaker is different, the speech sound constituting the speech voice changes significantly. , The same speech sound is mainly characterized by the spectral envelope and its temporal transition, and the stochastic nature of such a time series pattern of fluctuations can be accurately represented by the HMM. It is like this.

Therefore, in the present embodiment, as will be described later, the feature amount of the input speech voice and each keyword HM.
The keyword included in the uttered voice is recognized by calculating the degree of similarity with M and performing matching processing while considering unnecessary words.

In the present embodiment, this HMM takes the inverse of the data of the spectrum envelope showing the power at each frequency for every fixed time showing the feature quantity of the feature quantity pattern of each keyword or the logarithm of this power spectrum. A stochastic model having the cepstrum data obtained by performing the Fourier transform is shown.

In the case of performing speech recognition of a keyword included in a voice such as a spoken voice using such an HMM,
The recognized voice is divided at predetermined intervals, and the probability of changing from each divided state to the next state is calculated based on the matching process with the data of the keyword HMM stored in advance. By performing matching processing with a preset value as the unnecessary word similarity, which will be described later, the keywords included in the uttered voice and the unnecessary words are identified to determine the keywords to be recognized.

Specifically, in the present embodiment, the feature amount pattern of this HMM is obtained by comparing this feature amount pattern with the feature amount of each speech section of the utterance voice showing an arbitrary state divided into a certain time period. And a degree of similarity (corresponding to the keyword probability of the invention of the present application) indicating the degree of coincidence of the feature amount of each speech section, and the calculated similarity and each section of the uttered speech are assumed to be unnecessary words. A matching process described below is performed based on the voice feature amount of each voice section and a preset unnecessary word similarity, and the cumulative similarity (this is the cumulative similarity indicating the probability of any HMM connection, that is, the connection between the keyword and the unnecessary word). (Corresponding to the combination probability of the invention) is calculated, and the connection of the HMMs with the highest degree of similarity is recognized as the language of the spoken voice.

Next, the configuration of the speech recognition apparatus of this embodiment will be described with reference to FIG.

FIG. 2 is a block diagram showing the outline of the configuration of an embodiment of the word spotting voice recognition device according to the present invention.

The voice recognition device 100, as shown in FIG. 2, is a microphone 101 for inputting a speech voice to be recognized.
And a low pass filter (hereinafter LPF: Low Pass F
ilter) 102, an analog / digital conversion unit (hereinafter referred to as an A / D conversion unit) 103 that converts the audio signal output from the microphone 101 into a digital signal,
An input processing unit 104 that cuts out a voice signal of a uttered voice portion from a voice signal converted into a digital signal and divides into frames at preset time intervals, and a voice analysis unit that extracts a feature amount of the voice signal for each frame. 105, a keyword model database 106 in which a keyword HMM indicating a feature amount pattern of a plurality of keywords to be recognized is stored in advance, and a feature amount of each frame based on the extracted feature amount of each frame and the keyword HMM. A matching process described below is performed based on the similarity between the similarity calculating unit 107 that calculates the similarity and the calculated keyword HMM for each frame and the preset similarity of unnecessary words that do not form the keyword. Matching processing unit 1
08, and a determination unit 109 that determines a keyword included in the uttered voice based on the matching process.

The input processing unit 104 and the voice analysis unit 105 constitute the extracting means according to the present invention, and the keyword model database 106 constitutes the first storing means according to the present invention.

The similarity calculating section 107 constitutes the calculating means and the first obtaining means according to the present invention, and the matching processing section 108 constitutes the second storing means, the second obtaining means and the determining means, and The determination unit 109 constitutes the determination means according to the present invention.

Speech sound is input to the microphone 101, and this microphone 101
Generates a voice signal based on the input utterance voice, and
It is designed to be output to F102.

An audio signal generated by the microphone 101 is input to the LPF 102. The LPF 102 removes a high frequency component from the input audio signal and removes the high frequency component. The signal is output to the A / D conversion unit 103.

An audio signal from which the high frequency component has been removed by the LPF 102 is input to the A / D conversion unit 103. The A / D conversion unit 103 converts the input audio signal into an analog signal. Is converted into a digital signal, and the audio signal converted into the digital signal is input to the input processing unit 1.
It is designed to output to 04.

A voice signal converted into a digital signal is input to the input processing unit 104. The input processing unit 104 outputs a voice signal indicating the voice section of the uttered voice portion of the input digital signal. The voice signal of the cut voice segment is divided into frames at preset time intervals and is output to the voice analysis unit 105.

For example, the input processing section 104 is configured to divide one frame at time intervals of 10 ms to 20 ms.

A voice signal divided into frames is input to the voice analysis unit 105. The voice analysis unit 105 analyzes the voice signal for each input frame and also for each frame. The feature amount of the voice signal is extracted and output to the similarity calculation unit 107.

Specifically, the voice analysis unit 105 performs, for each frame, inverse Fourier transform by taking the information of the spectrum envelope showing the power at each frequency for every fixed time or the logarithm of this power spectrum. The information of the cepstrum thus obtained is extracted as a feature amount, and the extracted feature amount is vectorized and output to the similarity calculation unit 107.

The keyword model database 106 is
A keyword HMM indicating pattern data of the characteristic amount of the keyword to be recognized is stored in advance. The stored data of the plurality of keyword HMMs indicates the pattern of the characteristic amount of each of the plurality of recognition target words to be recognized.

For example, when used in a vehicle-mounted navigation device, the keyword model database 106
An HMM indicating a feature quantity pattern of a voice signal such as a destination name to which an automobile is heading, a current position name, a facility name such as a restaurant is stored in the.

In the present embodiment, the HMM indicating the feature amount pattern of each keyword is, as described above, the data of the spectrum envelope indicating the power at each frequency for each constant time or the inverse Fourier transform of the logarithm of this power spectrum. A probabilistic model having cepstrum data obtained by the conversion is shown.

Further, since a keyword is usually composed of a plurality of syllables or phonemes such as "current location" and "destination", in the present embodiment, one keyword HMM includes a plurality of syllables. The keyword calculating element 107 is configured to calculate the degree of similarity between the keyword calculating element 107 and the feature amount of each frame for each keyword forming element HMM.

As described above, each keyword H of the keywords to be recognized is stored in the keyword model database 106.
The MM, that is, the keyword component HMM is stored.

The vector feature amount for each frame is input to the similarity calculation unit 107. The similarity calculation unit 107 inputs the feature amount for each frame and the keyword model database 106. The similarity between the input frame and each HMM is calculated by comparing with the feature amount of the keyword HMM model stored in, and the calculated similarity is output to the matching processing unit 108. There is.

In this embodiment, the similarity calculation unit 107
Based on the feature amount of each frame and the feature amount of the HMM stored in the keyword model database 106, the probability that each frame indicates the HMM stored in the keyword model database 106 is calculated.

Specifically, the similarity calculator 107 calculates the output probability that each frame indicates each keyword constituent HMM, and the state transition from an arbitrary frame to the next frame is calculated from each keyword constituent HMM. A state transition probability indicating a state transition to the keyword component HMM or the unnecessary word component is calculated, and these probabilities are output to the matching processing unit 108 as the similarity.

The state transition probability includes the state transition probability indicating the state transition from each keyword constituent element HMM to its own keyword constituent element HMM.

Further, the similarity calculating section 107 outputs the output probabilities and the state transition probabilities calculated for each frame to the matching processing section 108 as the similarity of each frame.

The similarity calculation section 107 inputs the output probabilities and transition output probabilities for each frame to the matching processing section 108. The matching processing section 108 outputs the input outputs. A matching process is performed to calculate the cumulative similarity indicating the similarity of each combination of each keyword model HMM and the unnecessary word similarity, based on the probability, each transition output probability, and the unnecessary word similarity.
The calculated cumulative similarity is output to the determination unit 109.

Specifically, the matching processing unit 108
The output probability and the state transition probability indicating the unnecessary word similarity of the feature amount of the voice component of the frame and the feature amount of the voice component of the unnecessary word component when the frame is assumed to be the unnecessary word component, The cumulative similarity of all combinations of the keyword and the unnecessary word is stored in advance inside and the unnecessary word similarity and the similarity of each keyword calculated by the similarity calculating unit 107 are integrated for each frame. As described later, the cumulative similarity of 1 is calculated for each keyword, and the cumulative similarity in the case of no keyword is calculated.

Incidentally, such a matching processing unit 108
Details of the matching process performed in step 1 will be described later.

In this embodiment, when recognizing an unnecessary word that does not form a keyword in the uttered voice, the similarity between the feature amount of the uttered voice and the feature amount data of the unnecessary word is not calculated, A value set in advance is used as the unnecessary word similarity on the assumption that the feature quantities are similar.

For example, with respect to the unnecessary word similarity, the unnecessary word similarity is set to k in advance, a speech recognition experiment is actually performed while changing the value of k, and the highest performance of the changed values of k is obtained. A value calculated experimentally such as determining a good value as the unnecessary word similarity is used.

The same degree of similarity may be used before and after the keyword HMM, or different degrees of similarity may be set before and after the keyword HMM.

The unnecessary word similarity may be set in advance by the operation of the operation unit (not shown), or may be incorporated in advance when the speech recognition apparatus 100 is manufactured.

The determination unit 109 includes a matching processing unit 10
The cumulative similarity calculated for each keyword in 8 is input, and the input cumulative similarity is the word length of each keyword, that is, the temporal length of the keyword related to each cumulative similarity. The keyword having the highest degree of similarity among the normalized degrees of similarity is determined to be a keyword included in the uttered voice, and the keyword is output to the outside as a recognition result.

At this time, the determination unit 109 also determines the keyword by adding the cumulative similarity of only the unnecessary word similarity to the determination target, and the cumulative similarity of only the unnecessary word similarity is the highest. When the input cumulative similarity is the highest, it is determined that the speech voice does not include the keyword, and the determination result is output to the outside.

Next, the matching processing section 10 of the present embodiment.
The matching process performed in 8 will be described.

In the matching process of this embodiment,
The Viterbi algorithm is used, and the cumulative similarity of each combination of the keyword model and the preset unnecessary word similarity is calculated by this Viterbi algorithm.

This Viterbi algorithm is an algorithm for calculating the cumulative similarity based on the output probability of each state and the transition probability when transitioning from each state to another state. After calculating the cumulative similarity, the cumulative similarity is calculated. The combination for which the similarity is calculated is output.

In general, the Euclidean distance between the state indicated by the feature amount of each frame and the state of the feature amount indicated by the HMM is calculated, and the cumulative distance is calculated to calculate the cumulative similarity. It has become.

Specifically, the Viterbi algorithm is adapted to calculate a cumulative probability calculated along a path indicating a transition from an arbitrary state i to a next state j,
By performing this cumulative probability calculation, each path capable of state transition, that is, the connection and combination of HMMs is extracted.

In the present embodiment, based on the output probabilities and the state transition probabilities calculated by the similarity calculation unit 107, and the unnecessary word similarity set in advance, the first division of the input uttered voice is performed. From the frame to the last divided frame, each frame is sequentially applied with the output probability and the state transition probability of the keyword model, and the output probability and the state transition probability that are predetermined as unnecessary words, and the The cumulative probability from the first divided frame to the last divided frame by the combination is calculated, and the combination with the highest cumulative similarity calculated for each keyword model is given to the determination unit 109 one by one. It is designed to output.

For example, when the keywords to be recognized are “current position” and “destination” and the input speech voice is “um, current position”, the matching process of the present embodiment
The following processing is performed.

The unnecessary word is “Eh”, the unnecessary word similarity (output probability and state transition probability) is set in advance, and the keyword model database 10
In 6, the "current location" and "destination" are H for each syllable.
It is assumed that the MM is stored.

It is also assumed that the output probability and the state transition probability calculated by the similarity calculation section 107 have already been input to the matching processing section 108.

In such a case, in the present embodiment, the Viterbi algorithm is used to calculate the unnecessary word similarity and the cumulative similarity in all combinations for each of the "current location" keyword and the "destination" keyword. , The output probability and the state transition probability are calculated.

Specifically, when any uttered voice is input, "Genzaichi", "○ Genzaichi", "Genzaichi ○", "○ Genzaichi ○" (○ mark is unnecessary. (Fixed value of word similarity) In all combination similarities in each combination pattern, and in the keyword of the destination as well as the keyword of the current location, "Mokukuchi", "○ Mokukuchi",
All cumulative similarities in each pattern of “Mokukuichi ○” and “○ Mokukuchi ○” (indicated by “○” is unnecessary word similarity) are used as unnecessary word similarity, output probability, and state transition probability. It is based on the calculation.

The Viterbi algorithm is for each keyword model, in this case, for each of the "current position" and the "destination", sequentially from the first frame of the uttered speech, sequentially for each frame, and simultaneously for all cumulative similarities in all combination patterns. It is designed to calculate degrees.

Further, in the Viterbi algorithm, in the process of calculating the cumulative similarity of each combination for each keyword, a combination pattern with a low cumulative similarity is in the process of being calculated sequentially, and the uttered voice is not this combination pattern. Judgment is made and the calculation of the cumulative similarity is stopped.

Specifically, in the first divided frame, there is a similarity between the case where the HMM of "ge" which is the keyword constituent HMM of the keyword "current position" and the case where the unnecessary word constituent HMM is shown are similar. One of the degrees is added, but in this case, the one having the higher cumulative similarity is adapted to calculate the cumulative similarity of the next divided frame.
In the above case, the similarity of the unnecessary word component HMM is higher than the similarity of the HMM of “ge”, so “ge”
After that, the calculation processing of the cumulative similarity, that is, “Genzaichi ○” and “Genzaichi” is ended.

As a result, in such a matching process, the cumulative similarities in the keywords "current location" and "destination" are calculated one by one.

Next, the keyword recognition processing of this embodiment will be described with reference to FIG.

Note that FIG. 3 is a flowchart showing the operation of the keyword recognition processing of this embodiment.

First, when an instruction is input to each unit by an operation unit or a control unit (not shown) to start the keyword recognition process, and a speech voice is input to the microphone 101 (step S11), the LPF 102 and the A / D conversion unit 103 are input. Is input to the input processing unit 104 via the input processing unit 104, and the input processing unit 104 cuts out the voice signal of the uttered voice portion from the input voice signal (step S12), and performs frame division at preset time intervals. , For each frame, the audio signal is sequentially output to the audio analysis unit 105 from the first frame (step S13).

Next, this operation performs the following processing for each frame.

First, the control unit (not shown) determines whether or not the frame input to the voice analysis unit 105 is the final divided frame (step S14). When it is determined that the frame is the final divided frame, step S19. If it is determined that the frame is not the final divided frame, the following operation is performed.

First, the voice analysis unit 105 extracts the feature amount of the voice signal of the input frame and outputs the extracted feature amount of this frame to the similarity calculation unit 107 (step S15).

More specifically, the voice analysis unit 105 takes the inverse of the spectrum envelope information indicating the power at each frequency for each fixed time, or the logarithm of this power spectrum, based on the voice signal of each frame. The cepstrum information obtained by performing the Fourier transform is extracted as a feature amount, and the feature amount is vectorized and output to the similarity calculation unit 107.

Next, the similarity calculation unit 107 determines the feature quantity of the input frame and the keyword model database 1
06, the output probability and the state transition probability of the frame of each HMM model are calculated, and the output probability and the state transition probability are matched, as described above. Processing unit 108
(Step S16).

Next, the matching processing unit 108 outputs the output probabilities and the state transition probabilities calculated by the similarity calculating unit 107, and the unnecessary words which are preset and stored in the matching processing unit 108. Based on the degree of similarity, the above-described matching processing is performed to calculate the cumulative degree of similarity for each keyword (step S17).

Specifically, the matching processing unit 108
Each keyword HMM input to the cumulative similarity up to the last time
The similarity and unnecessary word similarity indicated by
Only the cumulative similarity having the highest cumulative similarity is calculated for each keyword type.

Then, the input control of the next frame is performed according to an instruction from the control unit (not shown) (step S18), and the process returns to step S14.

On the other hand, when the control unit (not shown) determines that the frame is the final divided frame, the highest cumulative similarity calculated for each keyword is output to the determination unit 109, and the determination unit 109 Normalization processing is performed on the word length of the cumulative similarity for each keyword (step S1).
9).

Finally, the determination unit 109 determines that the keyword having the highest similarity is the keyword included in the uttered voice, based on the normalized similarity for each keyword. Output to (step S2
0), this operation is ended.

As described above, according to this embodiment, the similarity indicating the ratio of the characteristics of the uttered voice feature and the keyword feature data is similar for each frame is calculated, and the calculated similarity and Since the keyword to be recognized included in the uttered voice can be determined based on the preset unnecessary word similarity, it is possible to calculate the characteristic of the uttered voice characteristic amount and the characteristic amount data of the unnecessary word without calculating the characteristic of the characteristic amount. By using the preset unnecessary word similarity, the keyword to be recognized can be determined.

Also, for each frame, the unnecessary word similarity and the calculated similarities are integrated to calculate the cumulative similarity in each combination of the unnecessary word similarity and the calculated similarities, and the cumulative similarity is calculated. Since it is possible to determine the keyword to be recognized included in the uttered voice based on the degree, it is possible to determine the keyword included in the uttered voice while considering each combination of the unnecessary word similarity and each calculated similarity. it can.

As a result, the keyword contained in the uttered voice can be accurately recognized easily and at high speed, and erroneous recognition can be prevented.

Further, in the present embodiment, when a plurality of keywords are recognized in one utterance voice, the keywords included in the utterance voice can be recognized more easily and at high speed, and erroneous recognition can be prevented. be able to.

For example, when recognizing two keywords, assuming a spoken language model 20 including keywords to be recognized using an HMM as shown in FIG. 4, it is based on the word length in each keyword model to be recognized. If the word length is normalized by using two words, the two keywords can be recognized at the same time.

That is, in the matching processing unit 108, instead of calculating the cumulative similarity for each keyword, the cumulative similarity is calculated for each combination of all the keywords stored in the keyword model database 106, and the determination is made. By adding the word lengths of the keywords and performing the normalization process in the unit 109, a plurality of keywords can be recognized at the same time, and the keywords included in the uttered voice can be easily and quickly recognized. It is possible to prevent erroneous recognition.

In this embodiment, the keyword recognition process is performed by the above-mentioned voice recognition device. However, the voice recognition device is provided with a computer and a recording medium, and a program for performing the above-mentioned keyword recognition process on this recording medium. May be stored and the keyword recognition processing program similar to the above may be performed by reading the keyword recognition processing program by this computer.

In this case, the recording medium is DV
It may be configured by a recording medium such as a D or a CD, and the voice recognition device 100 may be provided with a reading device that reads out a program from the recording medium.

[0129]

As described above, according to the first aspect of the invention, the characteristics of the feature amounts of the utterance voice feature amount and the feature amount data of the unnecessary words are not calculated, and the preset unnecessary By using the word probability, it is possible to determine the keyword to be recognized by distinguishing the unnecessary word from the keyword, so that the processing load when calculating the unnecessary word similarity can be reduced, and it is easy and fast. It is possible to recognize a keyword included in the uttered voice.

[Brief description of drawings]

FIG. 1 is a diagram showing a spoken language model showing a recognition network using an HMM.

FIG. 2 is a block diagram showing a schematic configuration of an embodiment of a word spotting voice recognition device according to the present invention.

FIG. 3 is a flowchart showing an operation of a keyword recognition process in one embodiment of the word spotting voice recognition device.

FIG. 4 is a diagram showing a spoken language model showing a recognition network using an HMM when recognizing the second keyword.

FIG. 5 is a diagram showing a spoken language model showing a filler model speech recognition network.

[Explanation of symbols]

100 ... Speech recognition device 101 ... Microphone 102 ... LPF 103 ... A / D conversion unit 104 ... Input processing unit (extraction means) 105 ... Speech analysis unit (extraction means) 106 ... Keyword model database (first storage means) 107 ... Similar Degree calculation unit (calculation unit, first acquisition unit) 108 ... Matching processing unit (determination unit, second storage unit, second acquisition unit) 109 ... Judgment unit (determination unit)

─────────────────────────────────────────────────── ─── Continuation of front page (51) Int.Cl. ⁷ Identification code FI theme code (reference) G10L 15/28

Claims (9)

[Claims] 1. A voice recognition device for recognizing a keyword included in a uttered voice, comprising: an extracting unit that extracts a uttered voice feature amount that is a feature amount of a voice component of the uttered voice by analyzing the uttered voice. A first storage unit that stores in advance keyword feature amount data indicating a feature amount of one or more voice components of the keyword; and the uttered voice feature amount extracted from at least a part of the voice section of the uttered voice. Calculation means for calculating each keyword probability indicating the probability that the uttered voice feature amount is the keyword based on the keyword feature amount data stored in the storage means; and a preset value for the uttered voice. Second storage means for storing as an unnecessary word probability indicating a probability that at least a part of the voice section is an unnecessary word that does not form the keyword; Determining means for determining the keywords to be recognized that is included in the uttered voice based on the keyword probabilities and the unnecessary word probability, the speech recognition apparatus characterized by comprising a. 2. The voice recognition apparatus according to claim 1, wherein the extraction unit analyzes the uttered voice for each preset unit time to extract the uttered voice feature amount information and stores the uttered voice feature amount information in advance. In the case where the unnecessary word probability stored in the means indicates the unnecessary word probability in the unit time, the calculating means calculates each of the keyword probabilities based on the extracted uttered voice feature amount for each unit time. The voice recognition device, wherein the determination unit determines the keyword to be recognized included in the uttered voice based on the calculated keyword probability and the unnecessary word probability in the unit time, in addition to the calculation. 3. The voice recognition device according to claim 2, wherein the determining unit determines the first probability based on the calculated keyword probabilities and unnecessary word probabilities in the unit time.
A combination probability indicating a probability of each combination of the keyword and the unnecessary word indicated by each of the keyword feature amount data stored in the storage means is calculated, and included in the uttered voice based on the combination probability. A voice recognition device characterized by determining the keyword to be recognized. 4. A voice recognition method for recognizing a keyword included in an uttered voice, comprising: extracting a uttered voice feature amount that is a feature amount of a voice component of the uttered voice by analyzing the uttered voice. A first acquisition processing step of previously acquiring keyword feature amount data indicating a feature amount of a voice component of one or more of the keywords; and the uttered voice feature extracted from at least a part of the voice section of the uttered voice. A calculation processing step of calculating each keyword probability indicating the probability that the uttered voice feature amount is the keyword based on the amount and the keyword feature amount data stored in the storage means; A second acquisition processing unit that acquires as an unnecessary word probability indicating a probability that at least a part of the speech section of the uttered voice is an unnecessary word that does not form the keyword. When the voice recognition method characterized by including: a determination step of determining the keyword to be recognized that is included in the uttered voice based on the keyword probabilities and the unnecessary words probability the calculated. 5. The voice recognition method according to claim 4, wherein in the extraction processing step, the uttered voice is analyzed for each preset unit time to extract the uttered voice feature amount information, and the calculation is performed. In the processing step, the keyword probabilities are calculated based on the extracted uttered voice feature amount for each unit time, and in the second acquisition processing step, the unnecessary word probability indicating the unnecessary word probability in the unit time. A voice recognition method characterized in that, in the determination processing step, the keyword to be recognized included in the uttered voice is determined based on the calculated keyword probability and the unnecessary word probability in the unit time. . 6. The speech recognition method according to claim 5, wherein, in the determination processing step, the acquired keyword feature amount is obtained based on the calculated keyword probability and the unnecessary word probability in the unit time. It is characterized in that a combination probability indicating a probability of each combination of each keyword indicated by the data and the unnecessary word is calculated, and the keyword to be recognized included in the uttered voice is determined based on the combination probability. Speech recognition method. 7. A voice recognition program for performing a voice recognition process of recognizing a keyword included in an uttered voice by a computer, wherein the computer analyzes the uttered voice to determine a feature amount of a voice component of the uttered voice. Extracting means for extracting the uttered speech feature amount, first acquisition means for obtaining in advance the keyword feature amount data indicating the feature amount of the voice component of one or more of the keywords, and at least a part of the voice section of the uttered voice. Calculating means for calculating each keyword probability indicating the probability that the utterance voice feature amount is the keyword based on the extracted utterance voice feature amount and the keyword feature amount data stored in the storage means, The determined value is an unnecessary word in which at least a part of the speech section of the uttered speech does not form the keyword. Second acquisition means for acquiring the probability of unnecessary words indicating a rate, and determination means for determining the keywords to be recognized included in the uttered voice based on the calculated keyword probabilities and the unnecessary word probabilities. Characteristic voice recognition program. 8. The voice recognition program according to claim 7, wherein the computer is configured to extract the uttered voice feature amount information by analyzing the uttered voice for each preset unit time, each unit. Calculation means for calculating the keyword probabilities based on the uttered voice feature quantity extracted for each time, second acquisition means for acquiring the unnecessary word probability in the unit time in advance, the calculated keyword probabilities and the unit time A voice recognition program, which functions as a determining unit that determines the keyword to be recognized, which is included in a speech voice, based on the unnecessary word probability in. 9. The voice recognition program according to claim 8, wherein the computer indicates the keyword feature amount data acquired in advance based on the calculated keyword probability and the unnecessary word probability in the unit time. A combination probability indicating a probability of each combination of each keyword and the unnecessary word is calculated, and based on the combination probability, it functions as a determination unit that determines the keyword to be recognized included in the uttered voice. Speech recognition method.