JP2003050595A - Device and method for speech recognition, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the precision of speech recognition by removing noise showing high likelihood as a result of speech recognition. SOLUTION: An inputted speech is sectioned into frames having specific time intervals and power component of the respective frames are found. A speech section is extracted according to the values of the power components and speech features of the respective speech sections are extracted. According to the speech features of the respective speech sections and a prepared hidden Markov phoneme model, the word having the highest likelihood among words contained in a word dictionary is extracted as a candidate. According to the acquired likelihood, maximum likelihood, and minimum likelihood of each frame, a likelihood reliability rate is found and compared with a previously set threshold. When the likelihood reliability rate exceeds the threshold, the vowel rate in the extracted speech section is found and then compared with a previously set threshold. When a vowel rate exceeds the threshold, the word as the candidate is outputted as a speech recognition result, but when the likelihood reliability rate or the vowel rate is less than each threshold, the word as the candidate is discarded.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and the like for recognizing speech of an unspecified speaker or a specified speaker.

[0002]

2. Description of the Related Art When performing voice recognition processing, irrelevant data such as noise is excluded as much as possible from PCM data obtained by converting the voice input by a speaker from a microphone into digital data, and only the voice uttered by the speaker is cut out. It is ideal to recognize and return the recognition result. Cut out irrelevant data,
Returning meaningless recognition results can cause confusion.

In order to construct a voice recognition system resistant to noise, in addition to improving the detection accuracy of the voice section, noise other than the voice uttered by the speaker is erroneously detected as the voice section. Even in this case, it is desirable to reject the erroneously detected section. If it is possible to reject a voice section that is erroneously detected from a source other than voice, it is possible to immediately return to the state of waiting for voice input, and it is possible to process as if nothing had happened.

Here, in the processing for detecting the voice section, generally, attention is paid to the power component of the voice, and it is judged that the portion where the power greatly increases is the voice start section. Then, a portion where the power is large does not continue for a long time and drops after a short time is considered as noise and is discarded.

[0005]

However, in such processing, it is not possible to reject noise having a power component similar to the voice uttered by the speaker, and it is determined that it is in the voice section. Will end up. Also, a voice such as "umu" that is unrelated to the content that the speaker originally wants to speak cannot be rejected only by detecting the voice section.

Therefore, all the portions detected as the voice section by the determination based on the power component are subjected to the voice recognition processing, and the probability that the recognition result is correct, that is, the likelihood of the recognition result is determined, and the voice recognition result is determined. Is output, or the detected voice section is rejected and the input waiting state is restored again. Here, as a method of obtaining the certainty of the recognition result, a method such as comparison with the result of continuous phoneme recognition, comparison with the difference with the likelihood of the garbage model, etc. is generally applied.

However, there are various kinds of noises generated in the real world, and some noises are judged to have a high likelihood of the recognition result. Especially in such an environment where a lot of noise is generated, the problem that the accuracy of voice recognition becomes low due to insufficient probability that the noise section can be rejected only by the conventional method described above. was there.

The present invention has been made in order to solve the above-mentioned problems of the prior art, and it is possible to reject noise showing a high likelihood as a result of the voice recognition, thereby improving the accuracy of the voice recognition. It is an object of the present invention to provide a voice recognition device and method capable of increasing the sound quality, and a program therefor.

[0009]

In order to achieve the above object, a speech recognition apparatus according to a first aspect of the present invention is a phoneme storing a phoneme model in which each phoneme included in a speech to be recognized is modeled. Model storage means, word dictionary storage means for storing a word dictionary in which phoneme pattern sequences of a plurality of types of words to be output as recognition results are stored, and input speech is referred to the phoneme model and the word dictionary. A candidate word extracting means for recognizing a voice by performing voice recognition, obtaining likelihoods of respective types of words registered in the word dictionary, and extracting candidate words output as a recognition result based on the obtained likelihoods; Ratio calculating means for calculating the ratio of vowels and consonants contained in the input voice, and ratio determination for determining whether the ratio of vowels and consonants calculated by the ratio calculating means is within a predetermined range When the stage, said ratio determining means determines to be in the predetermined range, characterized in that it comprises a result output means for outputting words the candidate word extracting means has extracted as a candidate as the speech recognition result.

In the above speech recognition apparatus, even if the input speech is speech-recognized and the candidate word is extracted from the candidate word extracting means, this is not output as it is as the final speech recognition result, and the input speech is further inputted. Is determined by the ratio of the vowels and the consonants included in. Even if the first candidate word with a high likelihood is extracted by voice recognition of noise, if the ratio of vowels and consonants is in a range that cannot be assumed as a voice spoken by a human, this is rejected and voiced. Since it is not necessary to output the recognition result, the accuracy of voice recognition can be increased.

In the above speech recognition apparatus, the ratio determining means can determine whether the ratio is in a predetermined range by comparing the ratio between the vowel and the consonant with a preset threshold value.

It is known that speech spoken by humans generally does not have an extremely low ratio of vowels contained therein. If the threshold value used for the comparison of the ratio determining means is set to a value that is not considered to be a voice uttered by a human, it is possible to judge that the ratio of vowels is extremely low as noise and reject it.

In the above speech recognition device, the word dictionary may be associated with each of the plurality of types of words and may further register a ratio range for use in the judgment of the ratio judgment means. In this case, the ratio determining means may determine whether the ratio between the vowel and the consonant is within the ratio range corresponding to the word extracted as a candidate by the candidate word extracting means.

The words include those with a high ratio of vowels and those with a low vowel ratio. By setting the ratio range for each word depending on whether the ratio of the original vowels is high or low, the number of cases in which noise can be rejected increases and it becomes possible to perform highly accurate speech recognition.

The voice recognition apparatus further comprises history registration means for registering a plurality of times of the history of the ratio of vowels and consonants calculated by the ratio calculation means for the same word input by the same speaker. Good. In this case, the ratio range registered in the word dictionary in association with at least one word may be calculated according to the history of the ratio registered in the history registration means.

It is known that when a specific speaker utters the same word several times, the ratio of vowels contained in the same word becomes almost the same. Therefore, the history of the ratio of vowels and consonants is registered especially for the voice recognition of a specific speaker,
By adjusting the ratio range based on the registered history, more accurate voice recognition can be performed.

In the above speech recognition apparatus, the candidate word extracting means may extract a plurality of word candidates ranked in order of increasing likelihood. In this case, the ratio calculating means calculates the ratio of vowels and consonants in order from the word having the highest likelihood, and the ratio determining means determines that the calculated ratio of vowels and consonants is not within a predetermined range. Occasionally, the ratio of vowels and consonants can be calculated for words in the next rank.

Even if the candidate word extracting means extracts the word having the highest likelihood as a candidate word, it may be rejected because of a low vowel ratio, and the speech recognition result cannot be output. However, even if the calculated likelihood is slightly low, the ratio of vowels may be high, and it may be the correct word. As described above, the plurality of words are ranked and the candidates are extracted, and the ratio of the vowel and the consonant is determined in order, which increases the possibility that the correct word can be output as the speech recognition result.

In the speech recognition apparatus, with respect to the words extracted as candidates by the candidate word extracting means, the likelihood reliability calculating means for calculating the reliability of the likelihood and the likelihood calculating by the likelihood reliability calculating means. It may further include a likelihood reliability determination means for determining whether or not the reliability of the degree exceeds a predetermined threshold. In this case, the result output means determines that the likelihood reliability determination means exceeds a predetermined threshold value, and when the ratio determination means determines that it is within a predetermined range,
The words extracted by the candidate word extracting means may be output as a voice recognition result.

It is empirically understood that even if a word is extracted as a result of speech recognition of noise, the reliability of the likelihood is low in this case. By adding such a configuration, it is possible to reject noise having a low reliability of likelihood, that is, noise having a likelihood lower than the treatment threshold, as noise, and thus it is possible to further improve the accuracy of voice recognition.

In the above speech recognition apparatus, when the candidate word extracting means extracts a plurality of word candidates ranked from the highest likelihood, the likelihood reliability calculating means determines the likelihood. The likelihood reliability is calculated in order from the highest word,
When it is determined by the likelihood reliability determination means that the reliability of the calculated likelihood does not exceed the predetermined threshold value,
The likelihood reliability can be calculated for the next ranked word.

Even if the candidate word extraction means extracts the word having the highest likelihood as a candidate word, the reliability of the likelihood becomes low and the word may be rejected, and the speech recognition result cannot be output. . However, even if the obtained likelihood is slightly low, the likelihood reliability may be high, and it may be the correct word. As described above, the plurality of words are ranked and the candidates are extracted, and the reliability of the likelihood is determined in order, which increases the possibility of outputting the correct word as the speech recognition result.

In the above speech recognition apparatus, the phoneme model may be a model in which each phoneme included in the speech to be recognized is modeled by a hidden Markov model. In this case, the candidate word extraction unit divides the input voice into a plurality of sections in units of a predetermined time, and obtains an acoustic feature vector for a voice section of the divided sections,
Continuous phoneme recognition is performed based on the acoustic feature vector of each section and the phoneme model, and a word having the maximum total likelihood in each frame can be extracted as a candidate.

In order to achieve the above object, the speech recognition method according to the second aspect of the present invention outputs an input speech as a phoneme model in which each phoneme included in a speech to be recognized is modeled and a recognition result. Speech recognition is performed by referring to a word dictionary in which phoneme pattern sequences of a plurality of types of words to be registered are respectively referred to, the likelihood of each type of word registered in the word dictionary is obtained, and based on the obtained likelihood. A candidate word extraction step of extracting a candidate word output as a recognition result, a ratio calculation step of calculating a ratio of vowels and consonants included in the input voice, and a vowel calculated in the ratio calculation step. A ratio determination step of determining whether the ratio with the consonant is within a predetermined range, and when the ratio determination step determines that the ratio is within the predetermined range, the candidate word extraction step is performed. Word extracted as candidates Te and characterized in that it comprises a result output step of outputting as a voice recognition result.

In the above speech recognition method, with respect to the words extracted as candidates in the candidate word extraction step, the likelihood reliability calculation step for calculating the reliability of the likelihood and the likelihood calculated in the likelihood reliability calculation step. And a likelihood reliability determining step of determining whether the reliability of the degree exceeds a predetermined threshold, wherein the result outputting step exceeds the likelihood reliability determining step predetermined threshold. When it is determined that it is within the predetermined range in the ratio determining step, the word extracted in the candidate word extracting step can be output as a voice recognition result.

In order to achieve the above object, the program according to the third aspect of the present invention outputs the input speech as a phoneme model in which each phoneme included in the speech to be recognized is modeled and a recognition result. Speech recognition is performed by referring to a word dictionary in which phoneme pattern sequences of a plurality of types of words to be registered are respectively referred to, the likelihood of each type of word registered in the word dictionary is obtained, and based on the obtained likelihood, Candidate word extraction means for extracting word candidates output as a recognition result,
Ratio calculating means for calculating the ratio of vowels and consonants contained in the input voice, ratio determining means for determining whether the ratio of vowels and consonants calculated by the ratio calculating means is within a predetermined range, and When the ratio determining means determines that the ratio is within a predetermined range, the computer device functions as result output means for outputting the word extracted as a candidate by the candidate word extracting means as a voice recognition result.

The program includes a likelihood reliability calculation means for calculating the reliability of the likelihood of a word extracted as a candidate by the candidate word extraction means, and a likelihood calculated by the likelihood reliability calculation means. The computer device can be made to function as a likelihood reliability determination means for determining whether or not the reliability of is greater than a predetermined threshold value, and in this case, the result output means is The word extracted by the candidate word extracting means may be output as a voice recognition result when it is determined that the threshold value is exceeded and the ratio determining means determines that the ratio is within a predetermined range.

[0028]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the accompanying drawings.

[First Embodiment] FIG. 1 is a block diagram showing the arrangement of a speech recognition apparatus according to this embodiment. As shown in the figure, this voice recognition device includes a voice input unit 1, an input voice storage unit 2, a power calculation unit 3, a voice section detection unit 4, a voice feature extraction unit 5, and a phoneme model storage unit 6. , The word dictionary storage unit 7, the likelihood calculation unit 8, and the verification unit 9
And a final determination unit 10 and an output unit 11.

The voice input unit 1 performs A / D conversion of voice (analog signal) input from an unspecified speaker, and a digital voice signal (eg, PCM) indicating this voice in time series.
Signal). The input voice storage unit 2 is the voice input unit 1
The digital audio signal converted by is stored.

The power calculator 3 converts the digital audio signal stored in the input audio storage 2 into a predetermined time interval (2.0).
Voice data is divided into a plurality of frames by a Hamming window such as a time window, and the voice data is cut out from each divided frame to obtain a power component.
The voice section detection unit 4 detects a frame in which the power component calculated by the power calculation unit 3 exceeds a predetermined threshold as a voice section. For example, a frame in which the power component falls below the threshold within 1 second is not set as a voice section. Speech feature extraction unit 5
Calculates the acoustic feature vector X (t) for each of the voice sections detected by the voice section detection unit 4 based on the power component calculated by the power calculation unit 3.

The phoneme model storage unit 6 stores a phoneme model obtained by modeling all the phonemes forming the speech to be recognized. As the phoneme model, for example, a hidden Markov model (HMM) as shown in FIG. 2 is applied. Figure 2
In the figure, only the phoneme "z" is shown, but in reality, HMMs for all phonemes are stored.

The word dictionary storage unit 7 stores a word dictionary in which phoneme pattern sequence information for each word is registered. This word dictionary is as shown in FIG. 3, for example, and z → e → n → for the word "Zenkoji".
Register a phoneme pattern sequence of k → o → u → j → i,
About the word "zentsuji" z → e
A phoneme pattern sequence of → n → ts →→ u → j → i is registered.

The likelihood calculator 8 calculates the acoustic feature vector X
The continuous phoneme recognition is performed for each frame with reference to (t) and the phoneme model based on the HMM, and the one having the maximum total likelihood is extracted as a first-rank candidate word. Likelihood calculation unit 8 calculates the minimum likelihood required for verification unit 9 to obtain likelihood reliability rate S, in addition to the likelihood (acquisition likelihood) required for extracting the first-ranked candidate word. Degree and maximum likelihood.

Next, the calculation of the likelihoods (acquisition likelihood, minimum likelihood and maximum likelihood) by the likelihood calculation section 8 will be described with reference to FIG. This is an example in the case of recognizing the word "zentsuji" by voice, and the ideal phoneme sequence is z ... z, e ... e, n ... n, t
s ... ts, u ... u, j ... j, i ... i.

In this figure, the vertical axis represents time (frame)
, And 80th to 79th frames are targets of likelihood calculation. The circles on the left side of the figure (just to the right of the frame number) represent the state of the HMM as a phoneme model, and the bold arrow indicates the state transition path when recognizing the word "Zentsuji." Shows. The value output in each frame for advancing this route is the acquisition likelihood.

The first phoneme is a phoneme that has the highest likelihood in the frame as a result of speech recognition of the speech signal stored in the input speech storage unit 2 for each frame, and the likelihood is the highest likelihood. Becomes In addition, the likelihood of the phoneme, which is the lowest as a result of speech recognition, becomes the lowest likelihood. The acquisition likelihood, the minimum likelihood, and the maximum likelihood in each frame are used by the next verification unit 9 to obtain the likelihood reliability rate S. Although a plurality of routes can be considered as the routes of the states from the 0th frame to the 79th frame, the one with the highest total of the acquisition likelihoods is extracted as the first candidate word.

The verification unit 9 selects the first candidate word extracted by the likelihood calculation unit 8 in order to improve the accuracy of speech recognition.
The likelihood reliability rate S is calculated by performing the following calculation.

[Formula 1] S = {Σ (acquisition likelihood-minimum likelihood) / (maximum likelihood-minimum likelihood)} / number of frames

When the likelihood reliability rate S is obtained, the verification unit 9
It is determined whether or not this is larger than a predetermined threshold value Smin. If it is less than or equal to the threshold value Smin, the first-ranked candidate word is rejected, and if it is greater than the threshold value, the final determination unit 10 further determines. Threshold Smin to be compared with likelihood reliability rate S
The value of is set to 90%, for example. This is set as a value for balancing the fact that some recognition result is obtained and that the recognition result is accurate.

In the final judgment unit 10, the verification unit 9 sets the threshold value Smin.
The vowel rate V of the first candidate word having the likelihood reliability rate S determined to be larger is calculated. The vowel rate V is determined by the voice section detection unit 4
Among the frames detected to be in the voice section, a, i,
A frame having u, e, o, and n phonemes is a vowel phoneme frame, and a frame having other phonemes is a consonant phoneme frame. For the total of the vowel phoneme frame and the consonant phoneme frame. It shows the ratio indicated by the frames of vowel phonemes.

Then, the final decision section 10 decides whether or not the obtained vowel rate V is larger than a preset threshold value Vmin. If it is larger than the threshold value Vmin, the first-ranked candidate word is passed to the output unit 11 as a correct word, and if it is equal to or smaller than the threshold value Vmin, the first-ranked candidate word is rejected. Here, the value of the threshold value Vmin that is compared with the vowel rate V is set to, for example, 15%. It is empirically known that the vowel rate of human speech in a language is relatively high, whereas, for example, as shown in FIG. Since it is extremely low, it is set to such a value.

The output unit 11 outputs the correct word passed from the final judgment unit 10. If the first candidate word is rejected and no correct answer is obtained, the output unit 11
The fact is output.

The voice recognition device can be realized by using a general-purpose computer such as a personal computer as a platform. For example, the voice input unit 1 and the output unit 11 are each a microphone and a display device connected to the general-purpose computer. Is realized by Input speech storage unit 2, phoneme model storage unit 6, word dictionary storage unit 7
Is realized by allocating a corresponding area in the memory. The phoneme model stored in the phoneme model storage unit 6 and the word dictionary stored in the word dictionary storage unit 7 are read in advance from an external device and stored in a memory before performing the speech recognition process. The power calculation unit 3, the voice section detection unit 4, the voice feature detection unit 5, the likelihood calculation unit 8, the verification unit 9, and the final determination unit 10 are CPUs (Central Processing).
Unit) executes the program stored in the memory.

Next, the operation will be described. FIG. 6 is a flowchart showing processing in the voice recognition device according to this embodiment.

First, the phoneme model, the word dictionary and the vowel rate table are read from the external storage device into the memory to construct the phoneme model storage unit 6 and the word dictionary storage unit 7 (step S101). Next, the value of the threshold value Smin that is compared with the likelihood reliability rate S and the value of the threshold value Vmin that is compared with the vowel rate V are set in the final determination unit 10 (step S102).
Then, the following voice recognition processing is performed (step S10
3) Then, the processing of this flowchart ends.

FIG. 7 is a flow chart showing the voice recognition process of step S103 in detail. First, a voice uttered by a speaker as a recognition target is input from the voice input unit 1, converted into a time-series digital voice signal, and stored in the input voice storage unit 2 (step S201). Next, the power calculation unit 3
Divides the digital audio signal stored in the input audio storage unit 2 into a plurality of frames at predetermined time intervals, and calculates the power component of each frame (step S202).

Next, the voice section detector 4 operates in step S2.
The power component of each frame calculated in 02 is compared with a predetermined threshold value, and a frame exceeding the threshold value is extracted as a voice section (step S203). Then, it is determined whether or not a voice segment necessary for extracting a candidate word is detected in this processing (step S204). If not enough voice sections have been detected, the process ends.
When the required number of voice sections are detected, the voice feature extraction unit 5 based on the power component of each frame calculated in step S202, the acoustic feature vector X (t) of the voice section extracted in step S203. Is calculated (step S205).

Next, the likelihood calculator 8 operates in step S205.
Of the word registered in the word dictionary stored in the word dictionary storage unit 7 based on the speech feature vector X (t) calculated in step 1 and the hidden Markov model as a phoneme model stored in the phoneme model storage unit 6. The word with the highest likelihood is extracted as the first candidate word (step S2).
06).

When the first candidate word is extracted, the verification unit 9
Performs the above-mentioned process to obtain the likelihood reliability rate S for the first-ranked candidate word (step S207). Then, the likelihood reliability rate S is compared with the threshold Smin set in step S102 to determine whether the likelihood reliability rate S exceeds the threshold Smin (step S208).
If the likelihood reliability rate S is less than or equal to the threshold value Smin, the extracted first candidate word is rejected (step S212), and the process ends.

If the likelihood reliability rate S exceeds the threshold value Smin, the first candidate word is passed to the final decision unit 10, and the final decision unit 10 makes a vowel sound in the voice section detected in step S203. The rate V is obtained (step S209). Then, the vowel rate V is set to the threshold value V set in step S102.
It is determined whether the vowel rate V exceeds the threshold value Vmin by comparing with min (step S210).

If the vowel rate V is less than or equal to the threshold value Vmin, the extracted first rank candidate word is rejected (step S21).
2), the process ends. On the other hand, when the vowel rate V exceeds the threshold value Vmin, the extracted first-ranked candidate word is passed to the output unit 11. Then, the output unit 11 outputs this first-ranked candidate word as a result of voice recognition (step S2).
11). Then, the process ends. Note that step S2
When the first candidate word is rejected in 12, a message indicating that is output from the output unit 11.

As described above, in the speech recognition apparatus according to this embodiment, the likelihood calculating section 8 extracts the first candidate word having the maximum likelihood, and then the verifying section 9 further.
The verification is performed based on the likelihood reliability rate S, and the final determination unit 10 performs the final determination based on the vowel rate V. Here, the likelihood calculation unit 8 only extracts the first-rank candidate word having the maximum likelihood, but when the frame detected by the voice section verification unit 4 as a voice section is a noise frame. Empirically, the value of the likelihood reliability rate S becomes low. Further, the vowel rate V of the voice uttered by the speaker does not become extremely low.

Therefore, when noise is recognized by voice,
If the likelihood reliability rate S becomes low or the vowel rate V becomes extremely low, even if the likelihood calculated by the likelihood calculation unit 8 is high, it can be rejected as noise. As a result, the rejection rate of noise can be increased, and speech recognition can be performed with higher accuracy than in the past.

[Second Embodiment] The speech recognition apparatus according to this embodiment has substantially the same configuration as that of the first embodiment, but is targeted for speech recognition of a specific speaker. However, the word dictionary stored in the word dictionary storage unit 7 is different from that of the first embodiment, and the word dictionary storage unit 7 further stores a vowel rate history table described later. There is. Further, the processing (program) executed for voice recognition is different from that of the first embodiment,
The function of the final determination unit 10 is different.

FIG. 8A is a diagram showing an example of a word dictionary stored in the word dictionary storage unit 7 in the voice recognition device according to this embodiment. In this word dictionary, in addition to the phoneme pattern sequence information for each word, a prescribed range (minimum value Vmin and maximum value Vma to be compared with the vowel rate V in the final determination unit 10 is added.
x) is registered. A predetermined default value is registered in the prescribed range in the initial state, but it can be changed based on the vowel rate registered in the vowel rate history table stored in the word dictionary storage unit 7.

FIG. 8B is a diagram showing a vowel rate history table stored in the word dictionary storage unit 7 in the voice recognition device according to this embodiment. The vowel rate history table is a table in which the history of the vowel rate V obtained by the final determination unit 10 is registered for each word registered in the word dictionary. In addition,
The final determination unit 10 performs the process of registering the vowel rate V in the vowel rate history table and the process of changing the specified range registered in the word dictionary.

Next, the operation will be described. FIG. 9 is a flowchart showing processing in the voice recognition device according to this embodiment.

First, as in the case of the first embodiment,
The phoneme model, the word dictionary, and the vowel rate history table are read from the external storage device into the memory, and the phoneme model storage unit 6
And the word dictionary storage unit 7 are constructed (step S30).
1). Next, the value of the threshold value Smin that is compared with the likelihood reliability rate S is set in the determination unit 10 (step S302). Then, the following voice recognition processing is performed (step S30
3) Then, the processing of this flowchart ends.

10 and 11 are flowcharts showing the voice recognition processing of step S303. Steps S401 to S409, S418 in this flowchart
The processing of each of steps S201 to S20 of FIG.
9, the same as the processing of S212.

When the vowel rate V is obtained in step S409, the final decision section 10 reads the minimum value Vmin and the maximum value Vmax of the specified range registered in the word dictionary in association with the first candidate word, It is determined whether the obtained vowel rate V is in a range larger than the minimum value Vmin and smaller than the maximum value Vmax, that is, within a specified range (step S41).
1). If it is not within the specified range, the process proceeds to step S418, and the first candidate word is rejected.

If it is within the specified range, the extracted first-rank candidate word is passed to the output unit 11, and the output unit 11 outputs this first-rank candidate word as a result of the speech recognition (step S).
411). Further, the final determination unit 10 substitutes the number of vowel rates already registered in the vowel rate history table for the output word into the variable n, and adds 1 to the value of the variable n (step S412). Then, step S409
The vowel rate V obtained in step S is registered in the vowel rate history table as the n-th vowel rate of the word (step S413).

Next, the final judgment section 10 judges whether or not the number n of vowel rates registered in the vowel rate history table for the word is equal to or larger than a predetermined number m (step S414). If the number n of vowel rates is not equal to or larger than the predetermined number m, the process is ended as it is.

On the other hand, if the number n of vowel rates is equal to or greater than the predetermined number m, the final judgment unit 10 obtains the variance p of the n vowel rates registered in the vowel rate history table for the word (step). S415). Next, it is determined whether the value of the variance p is smaller than the predetermined value x. Variance p
If the value of is greater than or equal to the value x, the process ends. When the value of the variance p is smaller than the value x, the minimum value Vmin of the specified range registered in the word dictionary for the word.
And the maximum value Vmax are added with the value of the variance p to reset the specified range (step S417). Then, the process ends.

As described above, in the voice recognition device according to the present embodiment, in the determination based on the vowel rate V, the specified range (minimum value Vmin and maximum value Vmax) registered for each word is used.
Is used. Therefore, as compared with the voice recognition apparatus shown in the first embodiment, noise can be rejected more accurately, and the accuracy of voice recognition is improved. Moreover, the value of the specified range can be changed according to the history of the vowel rate for the same word that the speaker has spoken in the past. Since this voice recognition device is intended for a specific speaker, the specified range can be made more appropriate by the history of vowel rates, and the accuracy of voice recognition can be improved.

[Modifications of the Embodiment] The present invention is not limited to the above-described first and second embodiments, and various modifications and applications are possible. Hereinafter, modifications of the above-described embodiment applicable to the present invention will be described.

In the above-described first embodiment, the threshold value Vmin compared with the vowel rate V is
Although the same value is used, as in the first embodiment, the threshold Vmin is registered in the word dictionary in association with each word, and the threshold Vmin associated with the first candidate word is
It may be compared with the obtained vowel rate V. here,
For words with a high vowel rate, such as "love," the threshold value Vmin is set to a high value, and "hisai"
For a word with a low vowel rate, the threshold value Vmin may be set low.

In the second embodiment described above, the specified range is registered in the word dictionary for all the words, but the unique specified range is registered only for some words, and for other words. Is a common specified range (VMIN, VM
AX) may be used. In this case, step S
In the process of 302, the common determination range (VMIN, VMAX) is set in the final determination unit 10. After obtaining the vowel rate V in step S409, it is determined whether or not a unique prescribed range is registered in the first-rank candidate word. If the original specified range is not registered, the process proceeds to step S410 as it is. If the original specified range is registered,
The common range (V
It is only necessary to rewrite from (MIN, VMAX) to an original specified range (Vmin, Vmax) and proceed to the processing of step S410.

In the second embodiment described above, the specified range (Vmin, Vmax) registered in the word dictionary is changed by adding the variance p of the vowel rate registered in the vowel rate history table. Was becoming. On the other hand, it is generally known that, if the same word is spoken by a specific speaker, the values of the vowel rates are almost the same. Therefore, as the number of vowel rates registered in the vowel rate history table increases, the specified range may be set narrower based on the calculation result.

In the first and second embodiments described above, when the verification unit 9 determines that the likelihood reliability rate S is larger than the threshold value Smin, the final determination unit 10 determines the vowel rate V. The decision was made. However, the vowel rate V may be determined without the processing of the verification unit 9, and the determination result may be used as the final result. Further, the vowel rate V is determined first, and when the vowel rate V is larger than the threshold value Vmin (in the second embodiment, the specified range (V
If it is determined to be within (min to Vmax)), the likelihood reliability rate S may be further compared with the threshold value Smin.

In the above first and second embodiments, the first
Although the likelihood reliability rate S and the vowel rate V of the rank candidate word are calculated based on all the frames, they may be calculated based on the frames at every two or more intervals. This reduces the calculation accuracy of the likelihood reliability rate S and / or the vowel rate V, but has the advantage that the processing time can be shortened.

In the above-described first and second embodiments, the likelihood calculating section 8 refers to the phoneme model by the HMM and calculates the likelihood by the continuous phoneme recognition for each frame. On the other hand, the likelihood may be calculated using a model in which the HMM is represented by a garbage model in which each vowel is mixed or each consonant is mixed. In this case, since voice recognition for each frame is not required, the amount of calculation is small, and it is suitable for application especially when the number of words registered in the word dictionary is small.

In the above first and second embodiments, the first
If the likelihood reliability rate S and the vowel rate V based on the rank candidate words are rejected, the speech recognition result cannot be obtained at that time. However, even if the likelihood is lower than that of the first-ranked candidate word, the word has a certain degree of likelihood (including the case where the word is the same as the first-ranked candidate word, only the path of the HMM state is different) If so, there is a possibility that the likelihood reliability rate S and the vowel rate V are not rejected in the determination, and that this is a correct word. Therefore, the following modifications may be added.

The likelihood calculating section 8 extracts, for example, three words having high likelihood as first, second, and third candidate words. If the first candidate word is rejected based on the likelihood reliability rate S or the vowel rate V, the likelihood reliability rate S and the vowel rate V of the second candidate word are sequentially obtained. If not rejected, the second candidate word is output from the output unit 11 as the voice recognition result. If the second candidate word is rejected based on the likelihood reliability rate S or the vowel rate V, the likelihood reliability rate S and the vowel rate V of the third candidate word are sequentially obtained. If not rejected, the output unit 11 outputs the third-ranked candidate word as the voice recognition result. If the third-ranked candidate word is also rejected, the fact that no voice recognition result has been obtained may be output from the output unit 11.

In the first and second embodiments described above, the output unit 11 is composed of, for example, a display device, and presents the voice recognition result (including the information indicating that it has been rejected) to the speaker. It was On the other hand, the output unit 11 may be connected to some electronic circuit and the electronic circuit may be operated according to the output result from the output unit 11.

In the first and second embodiments described above, FIG.
The program for executing the process of the flowchart shown in FIG. 7 or the process of the flowchart shown in FIGS. 9 to 11 has been described as being stored in the memory of the general-purpose computer. Then, the CPU executes the program stored in the memory to realize the power calculation unit 3, the voice section detection unit 4, the voice feature detection unit 5, the likelihood calculation unit 8, the verification unit 9, and the final determination unit 10. I was supposed to. On the other hand, the general-purpose computer is caused to execute the processing of the flowcharts of FIGS. 6 and 7 or the processing of the flowcharts of FIGS. 9 to 11, and the power calculation unit 3, the voice section detection unit 4, the voice feature extraction unit 5, the likelihood The program for realizing the power calculation unit 8, the verification unit 9, and the final determination unit 10 may be stored in a computer-readable recording medium such as a CD-ROM or a DVD-ROM and distributed.
Alternatively, these programs may be stored in a fixed disk device included in the Web server device and the Internet may be downloaded to a general-purpose computer.

[0076]

As described above, according to the present invention,
It is possible to reject even noise that shows a high likelihood as a result of voice recognition, and improve the accuracy of voice recognition.

[Brief description of drawings]

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

FIG. 2 is a diagram showing an example of a phoneme model stored in a phoneme model storage unit of FIG.

3 is a diagram showing an example of a word dictionary stored in a word dictionary storage unit of FIG.

FIG. 4 is a diagram illustrating a method of obtaining a likelihood reliability rate.

FIG. 5 is a diagram showing an actual measurement result of a sound of hitting a keyboard.

FIG. 6 is a flowchart showing processing in the voice recognition device according to the first embodiment of the present invention.

7 is a flowchart showing the voice recognition process of FIG. 6 in detail.

FIG. 8A is an example of a word dictionary stored in a word dictionary storage unit in the speech recognition apparatus according to the second embodiment of the present invention, and FIG. 8B is a diagram showing a vowel rate history table. Is.

FIG. 9 is a flowchart showing processing in the voice recognition device according to the second embodiment of the present invention.

10 is a flowchart showing the voice recognition process of FIG. 9 in detail.

11 is a flowchart showing the voice recognition process of FIG. 9 in detail.

[Explanation of symbols]

1 ... voice input unit, 2 ... input voice storage unit, 3 ... power calculation unit, 4 ... voice section detection unit, 5 ... voice feature extraction unit,
6 ... Phoneme model storage unit, 7 ... Word dictionary storage unit, 8 ...
Likelihood calculation unit, 9 ... Verification unit, 10 ... Final determination unit, 11 ...
・ Output section

Claims (12)

[Claims] 1. A phoneme model storage means for storing a phoneme model in which each phoneme included in a speech to be recognized is modeled, and a word for registering a phoneme pattern series of a plurality of types of words to be output as a recognition result. A word dictionary storage means storing a dictionary, and speech recognition of the input speech by referring to the phoneme model and the word dictionary, and the likelihood of each type of word registered in the word dictionary is obtained, and the obtained Candidate word extraction means for extracting word candidates output as a recognition result based on likelihood, ratio calculation means for calculating a ratio of vowels and consonants contained in the input speech, and the ratio calculation means When the ratio determining unit determines that the ratio of the calculated vowel and consonant is within a predetermined range, and the ratio determining unit determines that the ratio is within a predetermined range, There speech recognition apparatus characterized by comprising a result output means for outputting the words extracted as a candidate as the speech recognition result. 2. The ratio determining means compares the ratio between the vowel and consonant with a preset threshold value to determine whether or not the ratio is within a predetermined range. The voice recognition device described. 3. The word dictionary stores a ratio range to be used for the determination of the ratio determining means, in association with each of the plurality of types of words, and the ratio determining means determines that the vowel sounds The voice recognition device according to claim 1 or 2, wherein it is determined whether or not a ratio with a consonant is within a ratio range corresponding to a word extracted as a candidate by the candidate word extraction means. 4. For at least one word, at least one word further comprises history registration means for registering the history of the ratio of vowels and consonants calculated by the ratio calculation means for the same word input by the same speaker. The voice recognition device according to claim 3, wherein the ratio range registered in the word dictionary in association with each other is calculated according to the history of the ratios registered in the history registration means. 5. The candidate word extracting means extracts a plurality of word candidates ranked from the highest likelihood, and the ratio calculating means calculates the ratio of vowels and consonants in order from the highest likelihood word. And the ratio of the calculated vowels and consonants is not within a predetermined range, when the ratio determination means determines that the ratio of vowels and consonants is calculated for the word of the next rank. The voice recognition device according to claim 1, wherein 6. A likelihood reliability calculating means for calculating a reliability of likelihood of a word extracted as a candidate by the candidate word extracting means, and a reliability of likelihood calculated by the likelihood reliability calculating means. Further includes a likelihood reliability determination means for determining whether or not exceeds a predetermined threshold, the result output means determines that the likelihood reliability determination means exceeds a predetermined threshold, and the ratio determination means. 6. The speech recognition apparatus according to claim 1, wherein when it is determined that is within a predetermined range, the word extracted by the candidate word extraction unit is output as a speech recognition result. 7. The candidate word extracting means extracts a plurality of word candidates ranked from the highest likelihood, and the likelihood reliability calculating means calculates likelihoods in descending order of likelihood. The reliability is calculated, and when the likelihood reliability determination means determines that the calculated reliability of the likelihood does not exceed the predetermined threshold, the reliability of the likelihood of the word of the next rank is calculated. The voice recognition device according to claim 6, wherein 8. The phoneme model is a model in which each phoneme included in a speech to be recognized is modeled by a hidden Markov model, and the candidate word extracting means sets the input speech in units of a predetermined time. Dividing into a plurality of sections, obtaining an acoustic feature vector for a voice section of the divided sections, performing continuous phoneme recognition based on the acoustic feature vector of each section and the phoneme model, the likelihood of each frame The speech recognition device according to claim 1, wherein a word having a maximum total is extracted as a candidate. 9. A word dictionary in which a phoneme model obtained by modeling each phoneme included in a voice to be recognized and a phoneme pattern series of a plurality of types of words to be output as a recognition result are registered in a word dictionary. Voice recognition by referring to
Obtaining the likelihood of each type of word registered in the word dictionary,
A candidate word extraction step of extracting a candidate word output as a recognition result based on the obtained likelihood; a ratio calculation step of calculating a ratio of vowels and consonants included in the input speech, A ratio determination step of determining whether the ratio of vowels and consonants calculated in the ratio calculation step is within a predetermined range, and when it is determined that the ratio is within a predetermined range in the ratio determination step, in the candidate word extraction step And a result output step of outputting a word extracted as a candidate as a voice recognition result. 10. A likelihood reliability calculation step of calculating a reliability of likelihood of a word extracted as a candidate in the candidate word extraction step, and a reliability of likelihood calculated in the likelihood reliability calculation step. Further includes a likelihood reliability determination step of determining whether or not exceeds a predetermined threshold, the result output step determines that the likelihood reliability determination step exceeds a predetermined threshold, and in the ratio determination step 10. The voice recognition method according to claim 9, wherein the word extracted in the candidate word extraction step is output as a voice recognition result when it is determined that the word is within a predetermined range. 11. A word dictionary in which a phoneme model obtained by modeling each phoneme included in a voice to be recognized and a phoneme pattern series of a plurality of types of words to be output as a recognition result are registered. Candidate word extraction means for performing speech recognition with reference to obtain likelihood of each type of word registered in the word dictionary, and extracting a candidate of a word output as a recognition result based on the obtained likelihood. Ratio calculating means for calculating the ratio of vowels and consonants contained in the input voice, ratio determining means for determining whether the ratio of vowels and consonants calculated by the ratio calculating means is within a predetermined range, and When the ratio determining unit determines that the ratio is within a predetermined range, the computer apparatus is provided as a result output unit that outputs a word extracted as a candidate by the candidate word extracting unit as a voice recognition result. Program for the functioning of the. 12. A likelihood reliability calculation means for calculating a reliability of likelihood of a word extracted as a candidate by the candidate word extraction means, and a reliability of likelihood calculated by the likelihood reliability calculation means. The computer device functions as a likelihood reliability determination means for determining whether the degree exceeds a predetermined threshold, the result output means determines that the likelihood reliability determination means exceeds a predetermined threshold, and The program according to claim 11, wherein when the ratio determining unit determines that the ratio is within a predetermined range, the word extracted by the candidate word extracting unit is output as a voice recognition result.