JP2003044079A - Device and method for recognizing voice, recording medium, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the recognition accuracy with respect to a background noise. SOLUTION: An audio signal inputted from a microphone 1 is supplied through an A/D converting part 2 to a voice synthesizing part 34. In the voice synthesizing part 34, a noise signal stored in a noise recording part 33 is read and superimposed on the inputted audio signal and the superimposed audio signal is supplied to an acoustic analyzing part 3. In the acoustic analyzing part 3, acoustic analyzing processing is applied to the inputted audio signal, and a feature vector as a feature amount is extracted and supplied to a register part 11. In the register part 11, an acoustic model is read out of an acoustic model database 5, the acoustic model is connected on the basis of an acoustic model network 12, and the arrangement of the acoustic models of the highest score (phoneme sequence) is determined to the inputted feature amount and registered at a language model database 6 as sounding information of a correspondent word.

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 method, a recording medium, and a program, and more particularly, to a voice recognition device and method, a recording medium, and a recording medium that can improve the recognition accuracy for background noise. Regarding the program.

[0002]

2. Description of the Related Art Recently, a voice recognition device has been used in many systems as a man-machine interface or the like.

FIG. 1 shows the configuration of an example of such a voice recognition device.

The voice uttered by the user is the microphone 1
Is input to the microphone 1 and the input voice is
It is converted into a voice signal as an electric signal. This audio signal is supplied to the AD (Analog Digital) converter 2. AD
The converter 2 samples and quantizes a voice signal which is an analog signal from the microphone 1 and converts it into voice data which is a digital signal. This voice data is supplied to the acoustic analysis unit 3.

The acoustic analysis unit 3 performs acoustic analysis processing on the audio data from the AD conversion unit 2 for each appropriate frame (at a minute time interval), whereby, for example, MFC is performed.
A feature vector as a feature amount such as C (Mel Frequency Cepstrum Coefficient) is extracted and supplied to the recognition unit 4.
In addition, in the acoustic analysis unit 3, it is possible to extract other features such as a spectrum, a linear prediction coefficient, a cepstrum coefficient, and a line spectrum pair. This analysis includes linear prediction analysis (LPC), fast Fourier transform (FF)
T), a band pass filter (BPF), etc. are used.

The recognizing unit 4 refers to the acoustic model database 5 and the language model database 6 composed of word dictionary information and grammatical rule information as necessary using the feature quantity sequence from the acoustic analysis unit 3, The voice (input voice) input to the microphone 1 is, for example, continuously distributed HM.
Speech recognition based on the M method and the like.

The acoustic model database 5 stores acoustic models representing acoustic features such as individual phonemes and syllables in the language of speech to be recognized. Here, since speech recognition is performed based on the continuous distribution HMM method, for example, an HMM (Hidden Markov Model) is used as the acoustic model. In the language model database 6, for each word (vocabulary) to be recognized, how the word dictionary information describing the pronunciation (phonological information) and each word registered in the word dictionary information are linked. It stores grammatical rule information (language model) describing whether it is (connected). Here, as the grammar rules, for example, context-free grammar (CFG) and statistical word chain probability (N-gr) are used.
am) etc. are used.

The recognition section 4 refers to the word dictionary of the language model database 6 and connects the acoustic models stored in the acoustic model database 5 to form an acoustic model of a word (word model). Further, the recognition unit 4 connects some word models by referring to the grammar rule information stored in the language model database 6, and uses the word models thus connected to connect the word models based on the feature amount. The voice input to the microphone 1 is recognized by the continuous distribution HMM method. That is, the recognition unit 4 detects the word model sequence having the highest score (likelihood) at which the time-series feature amount output from the acoustic analysis unit 3 is detected, and the word string corresponding to the word model sequence is detected. , Output as a voice recognition result.

That is, the recognizing unit 4 accumulates the appearance probabilities of the respective feature quantities for the word string corresponding to the connected word model, and uses the cumulative value as a score to determine the word string having the highest score. Output as a recognition result.

Specifically, for example, from among the phonemes or syllables of the acoustic model stored in the acoustic model database 5, the Japanese words "a", "i", "u", "e",
If you use a pseudonym acoustic model with "o", "ka" ..., "n" as a unit, you can connect them to create "yes",
You can compose various words such as "No", "Good morning", and "What are you doing now?" And
With respect to these words, it becomes possible to calculate a score indicating the degree of similarity with the input feature amount.

The information connecting the acoustic models is the word dictionary information and the grammar rule information of the language model database 6. The word dictionary information is information that gives how to connect acoustic models in order to form each word to be recognized. Grammar rule information is information that gives how to connect words to each other. For example, "(number)
When dealing with the sentence "from time to (number) hour", first,
The numbers "0", "1" ..., "24" and the words "hour", "from" and "to" are read respectively. By including the kana as word dictionary information, the connection relation of the acoustic models in units of kana is given. Next, "(number)" + "hour" +
The connection relation of words is given by having the rule “from” + “(number)” + “hour” + “up” as grammar rule information. By combining these word dictionary information and grammatical rule information, it is possible to calculate the similarity between each sentence and the input feature amount, such as "1 to 2" or "2 to 5". It is possible to output the one with the highest score as the recognition result.

Therefore, this speech recognition apparatus can recognize various words only by changing the word dictionary information or the grammatical rule information by using a minute unit such as a phoneme or a syllable as an acoustic model.

However, in the case of constructing such a speech recognition apparatus, information about connection relations of acoustic models (hereinafter referred to as pronunciation information), such as phonetic kana, for each word is used as word dictionary information and a language model is created. It is necessary to set it in the database 6 in advance. For example, when one's own name is not registered in the language model database 6, it is sufficient to input a phonetic kana by using a keyboard or the like, but it is troublesome to input the kana. Furthermore, in addition to widely used units such as Kana, when units such as phonemes and phonetic symbols that are rarely used in daily life are used as acoustic models, the connection relation (pronunciation information) can be calculated without prior knowledge. Entering is extremely difficult.

Therefore, there is a method of using a phoneme typewriter for the problem of how to input pronunciation information for a new word. The phoneme typewriter is for performing phoneme recognition on input speech to estimate a corresponding phoneme sequence (pronunciation information). As a unit of recognition, various units such as syllables may be used in addition to phonemes.

Next, acquisition of pronunciation information corresponding to a new word not included in the word dictionary information and acquisition of new pronunciation information for a word included in the word dictionary information will be described.

FIG. 2 shows another example of the configuration of a voice recognition device having a registration function using a phoneme typewriter. Note that, in the drawing, the portions corresponding to those in FIG. 1 are denoted by the same reference numerals and the description will be repeated, and therefore the description thereof will be omitted below as appropriate.

As a unit of the acoustic model database 5, a minute unit such as a phoneme or a syllable, for example, a phoneme HMM as shown in FIG. 3A in which a vowel and a consonant are used. “Sil” in FIG. 3A indicates an HMM that models a voiceless portion. The language model database 6 is composed of word dictionary information and grammatical rule information, and information (pronunciation information) about a phoneme HMM connection method is registered corresponding to each word included in the word dictionary information. For example, as shown in FIG. 3B, pronunciation information “hai” is registered corresponding to the word “yes”.

The acoustic analysis unit 3 extracts a feature amount from the input voice input signal and supplies it to the registration unit 11. Registration department 1
1 performs a voice recognition process using the acoustic model database 5 and the acoustic model network 12. The voice recognition processing is performed by connecting the acoustic models based on the acoustic model network 12 and determining the arrangement of the acoustic models having the highest score for the input feature amount.

As shown in FIG. 4, the acoustic model network 12 includes acoustic models "a", "i", "u" ...
The state transition network has “N” and “sil” as nodes, and is configured as a network capable of generating an arbitrary sequence of acoustic models, that is, an arbitrary phoneme sequence (pronunciation information). For example, "hai" means "ST
From “ART”, go through “h” via branch point 21, return from branch point 22 to branch point 21, pass through “a”, then branch point 2
It is generated from the state transition from 2 to the branch point 21, passes through “i” and the branch point 22, and reaches “END”.

The calculation of the score is obtained by connecting the phoneme HMMs based on the speech model network 12 and accumulating the probability values of outputting the input feature amount on the network. For example, Viterbi
An algorithm accumulation method is used. This makes it possible to determine the state transition sequence having the highest cumulative value for one feature amount sequence. That is, it is possible to obtain the phoneme sequence (pronunciation information) that has the highest score in the sequence of all phoneme HMMs.

The pronunciation information obtained here is supplied to the language model database 6 and newly registered in the word dictionary information of the language model database 6 as the pronunciation information about the corresponding word. In addition, a plurality of pronunciation information may be given to one word.

As described above, the language model database 6
The word dictionary information of is updated as appropriate. Then, in the voice recognition process, the updated word dictionary information is used.
Therefore, by acquiring the pronunciation information, it becomes possible to add a new word to the system or correct the pronunciation information about the word already held inside.

In this way, when the pronunciation information is correctly given to a certain word, a high score is usually given to the corresponding voice by connecting the acoustic model according to the pronunciation information. You can

[0024]

However, when background noise or the like is added to the input voice, the feature amount obtained by the acoustic analysis is changed, so that a high score is not always given.

For example, when a phoneme sequence is estimated using a phoneme typewriter based on a voice signal uttered in a quiet environment, the voice sequence does not match a voice signal added with background noise. . That is, there was a problem that the recognition rate was lowered.

The present invention has been made in view of such a situation, and makes it possible to improve the recognition accuracy for background noise.

[0027]

The speech recognition apparatus of the present invention is composed of an acquisition unit for acquiring background noise, a combination unit for combining the input noise with the background noise acquired by the acquisition unit, and a combination unit. Acoustic analysis of the synthetic speech, analysis means for extracting the feature amount of the synthetic voice, estimating means for estimating pronunciation information based on the feature amount extracted by the analyzing means, pronunciation information estimated by the estimating means, And a registration unit that registers as pronunciation information of the corresponding word.

The registration means can register a plurality of pronunciation information for a word.

It is possible to further include a matching means for performing matching processing based on the pronunciation information registered by the registration means.

The speech recognition method of the present invention includes an acquisition step of acquiring background noise, a synthesizing step of synthesizing background noise with input speech by the processing of the obtaining step, and an acoustic analysis of the synthesized speech synthesized by the processing of the synthesizing step. Then, the analysis step of extracting the feature amount of the synthesized speech, the estimation step of estimating pronunciation information based on the feature amount extracted by the processing of the analysis step, and the pronunciation information estimated by the processing of the estimation step are associated with each other. And a registration step of registering as pronunciation information of a word to be performed.

The program of the recording medium according to the present invention includes an acquisition step of acquiring background noise, a synthesizing step of synthesizing the background noise with the input voice by the processing of the obtaining step, and a synthesized voice synthesized by the processing of the synthesizing step. An analysis step of analyzing and extracting the feature amount of the synthesized voice;
An estimation step of estimating pronunciation information based on the feature amount extracted by the processing of the analysis step; and a registration step of registering the pronunciation information estimated by the processing of the estimation step as pronunciation information of the corresponding word. Characterize.

According to the program of the present invention, a computer for a voice recognition device that performs voice recognition processing for recognizing an input voice synthesizes background noise by an acquisition step of acquiring background noise and an input voice processing in the acquisition step. A synthesis step, acoustic analysis is performed on the synthesized speech synthesized by the processing of the synthesis step, and an analysis step for extracting the feature amount of the synthesized speech, and estimation of pronunciation information based on the feature amount extracted by the processing of the analysis step It is characterized in that the estimation step and the registration step of registering the pronunciation information estimated by the processing of the estimation step as the pronunciation information of the corresponding word are performed.

In the speech recognition apparatus and method, the recording medium, and the program of the present invention, background noise is synthesized with the input speech, the synthesized speech is acoustically analyzed, and the characteristic amount of the synthesized speech is extracted to obtain the characteristic amount. The pronunciation information estimated based on this is registered as the pronunciation information of the corresponding word.

[0034]

FIG. 5 shows an example of the configuration of a voice recognition device to which the present invention is applied. Note that, in the figure, portions corresponding to those in FIG. 1 and FIG. 2 are denoted by the same reference numerals, and the description will be repeated, so the description thereof will be omitted below as appropriate.

The control unit 32 controls the AD conversion unit 2 based on the user's instruction from the input unit 31, and the AD conversion unit 2
The audio digital data input to is output to the acoustic analysis unit 3 or the audio synthesis unit 34.

When the voice input signal is input from the AD conversion unit 2, the voice synthesis unit 34 reads the noise signal stored in the noise storage unit 33, superimposes it on the input voice input signal, and the acoustic analysis unit 3 Supply to.

When the voice input signal is input from the voice synthesizing unit 34, the acoustic analysis unit 3 extracts the feature amount and supplies it to the recognition unit 4 or the registration unit 11.

The noise signal stored in the voice synthesizer 34 is stored in the noise memory 33. For example,
When recognizing voice in a moving vehicle, only noise during driving is recorded and stored in advance, or noise or fan noise corresponding to various driving conditions is recorded and stored even in the same vehicle. . That is, this noise signal is estimated and stored to some extent in advance.

The pronunciation information registration process of the voice recognition device will be described with reference to the flowchart of FIG.

The voice uttered by the user is the microphone 1
Is input to the microphone 1 and the input voice is
It is converted into a voice signal as an electric signal. In this example,
The registration is performed based on the utterance in the stopped vehicle, and the voice recognition is performed while the vehicle is running. Therefore, the noise recording unit 33 stores the traveling noise (noise signal) in advance.

In step S1, the AD conversion section 2
A voice signal is input via the microphone 1.

In step S2, the control unit 32, based on the user's instruction from the input unit 31, the AD conversion unit 2
When it is determined whether or not to superimpose a noise signal on the voice signal input to, and when it is determined to superimpose the noise signal on the voice signal,
The AD converter 2 is controlled to supply the audio signal to the audio synthesizer 34.

In step S3, the voice synthesizer 34
Reads the noise signal stored in the noise recording unit 33 and superimposes it on the audio signal input from the AD conversion unit 2.
Then, the voice synthesis unit 34 supplies the superimposed voice signal to the acoustic analysis unit 3.

When it is determined in step S2 that the noise signal is not superimposed on the voice signal, the AD conversion unit 2 supplies the voice signal to the acoustic analysis unit 3. At this time, the process of superimposing the noise signal in step S3 is skipped.

In step S4, the acoustic analysis unit 3
An acoustic analysis process is performed for each appropriate frame of the input voice signal (at a minute time interval), whereby a feature vector as a feature amount is extracted and supplied to the registration unit 11.

In step S5, the registration unit 11 reads the acoustic model from the acoustic model database 5,
The acoustic models are connected based on the acoustic model network 12. In step S6, the registration unit 11 determines the arrangement (pronunciation information) of the acoustic model having the highest score from the acoustic models connected to the feature amount input from the acoustic analysis unit 3.

In step S7, the registration unit 11 registers the determined pronunciation information in the language model database 6 as the pronunciation information of the corresponding word.

In the above processing, it is possible to register a plurality of pronunciation information for one word.
It is also possible to generate and register two types of pronunciation information, a voice signal in which a noise signal is not superimposed and a voice signal in which a noise signal is superimposed.

The pronunciation information registration process using the noise storage unit 33 is performed by the language model database 6 as described above.
This is not only performed for new words that are not included in the word dictionary information, but is similarly performed for words that are already included in the word dictionary information. Thereby, it is possible to register the pronunciation information in consideration of the background noise.

The word dictionary information of the language model database 6 updated in this way will be used in the speech recognition process described below.

The voice recognition processing of the voice recognition device will be described with reference to the flowchart of FIG.

In step S21, the AD conversion unit 2
Supplies the audio signal input via the microphone 1 to the acoustic analysis unit 3.

In step S22, the acoustic analysis unit 3
Performs acoustic analysis processing on the input voice signal for each appropriate frame (at minute time intervals), thereby extracting a feature vector as a feature amount, and recognizing unit 4
Supply to.

In step S23, the recognizing unit 4 refers to the word dictionary information of the language model database 6 based on the input feature amount series, and connects the acoustic models stored in the acoustic model database 5 to each other. , A word acoustic model (word model).

In step S24, the recognizing unit 4 accumulates the appearance probabilities of the respective feature quantities with respect to the word string corresponding to the connected acoustic model, and uses the cumulative value as a score to find the word string having the highest score. , Output as a voice recognition result.

As described above, since the pronunciation information in consideration of the noise signal obtained in the process of FIG. 6 is registered in the word dictionary information of the language model database 6, in the presence of background noise, the voice When recognizing, it is possible to suppress a decrease in recognition rate.

Actually, the experimental results of the evaluation of the voice recognition performed using the voice recognition device of the present invention will be described.

In a quiet environment, data obtained by uttering 100 predetermined words three times by five men and five women are recorded. The first two data were recorded on the same day and were used to register pronunciation information. The last piece of data was recorded one month later and was used for the evaluation of speech recognition.

In this example, the acoustic model used is a model in which a phoneme HMM in three states is created with respect to 29 types of phonemes, taking into consideration the dependence on the preceding and following phoneme environments.
Moreover, MFCC was used for the acoustic analysis of the acoustic analysis unit 3. Furthermore, the input speech signal was subjected to noise removal called spectral subtraction.

FIG. 8A shows an experimental result of evaluation of speech recognition performed using the word dictionary information of the language model database 6 composed of 100 words. The evaluation target is the data of the third utterance, which was recorded in a quiet environment. The recognition rate shows the average value of 10 persons.

The recognition rate of "phonetic alphabet" indicates the recognition rate when pronunciation information is determined from the phonetic alphabet, and its value is 99.3.
It was 0%. This is the result of speech recognition using the pronunciation information (recognition method described with reference to FIG. 1) of the phonetic model of the phonetic transcription in the language model database 6. That is, the registration unit 11 and the voice synthesis unit 34 (noise storage unit 33) are not used.

The "single-voiced" recognition rate indicates the recognition rate when only one pronunciation information is registered from the first vocalization, and the value is 99.10%, and the "double-voiced" recognition rate. Is 1
The recognition rate when the two pieces of pronunciation information obtained from the second utterance and the pronunciation information obtained from the second utterance were registered was shown, and the value was 99.50%. These are the results of processing of speech recognition using the pronunciation information (recognition method described with reference to FIG. 2) registered using the phoneme typewriter. That is, the voice synthesis unit 34 (noise storage unit 33)
Is not used.

From the above, 2 using the phoneme typewriter
By registering the two types of pronunciation information obtained from the utterances of the times, it can be seen that the recognition rate is almost the same as that when the pronunciation information is determined from the reading kana.

In the following, when registering the pronunciation information, the first two
The description will be made assuming that the two types of pronunciation information obtained from the vocalizations are registered.

FIG. 8B shows an experimental result of evaluation of voice recognition performed using the word dictionary information of the language database 6 composed of 100 words. The evaluation target is the data of the third utterance and the running noise of the car is superimposed. The running noise of this car is 7 kinds of in-vehicle noise including difference in vehicle type, difference in running speed, difference in road condition, etc. Recognition is being evaluated. Therefore, the recognition rate indicates the average value of 10 persons in 8 kinds of environments.

The recognition rate of "Yomikana" is 92.34%, the recognition rate of "conventional" is 92.15%, the recognition rate of "Invention 1" is 94.88%, and "Invention 2". The recognition rate was 95.22%.

"Yomi kana" indicates the case where the pronunciation information is determined from the phonetic kana, and "conventional" indicates the case where the pronunciation information is registered using only the phoneme typewriter. Further, "present invention 1" shows a case where, in addition to the pronunciation information registered using the phoneme typewriter, the pronunciation information in consideration of the background noise stored in the noise storage unit 33 as described above is registered. , "Present invention 2" shows a case where, in addition to the pronunciation information of "present invention 1", pronunciation information obtained from a phonetic kana is also used. In both cases using a phoneme typewriter or background noise, two vocalizations are used to obtain pronunciation information. Therefore, for example, as shown in FIG. 9, the number of pronunciation information for one word is one of “be N ch i” for “phonetic kana”,
"Conventional" is two (two of the present invention 1), "Invention 1" is "hbemu ch ii", "pr deu ch i", "b"
"eri N gi" and "pe N ch ij" are four, and "present invention 2" is five (reading kana + present invention 1).

As described above, in the environment including the in-vehicle noise, as compared with the previous result, in the case of using "Yomikana",
The recognition rate has dropped from 99.30% to 92.34%. Similarly, the “conventional” method, which does not consider background noise, also has a recognition rate of 92.15%.

On the other hand, when “Invention 1” in which the background noise is taken into consideration is used, the performance is improved to 94.88%, and when “Invention 2” is used, the performance is improved to 95.22%.

In particular, the "present invention 2" is the one in which pronunciation information in consideration of background noise is added to the word dictionary information of the language model database 6 in which pronunciation information is predetermined from the phonetic alphabet. As a result, it can be seen that the recognition rate can be improved by applying the present invention even to a word registered in advance in the word dictionary information.

FIG. 8C shows 50 including 100 words for evaluation.
It is an experimental result of the evaluation of the speech recognition performed using the word dictionary information of the language model database 6 composed of 75 words. The evaluation target is the data of the third utterance, and FIG.
As in the case of, the evaluation of speech recognition in eight types of noise environments is being performed. Therefore, the recognition rate indicates the average value of 10 persons in 8 kinds of environments.

The recognition rate of "Yomikana" was 71.28%, and the recognition rate of "present invention" was 86.80%.

In the "present invention", pronunciation information in which background noise is taken into consideration for 100 words for evaluation is added to the word dictionary information of the language model database 6 of 5075 words whose pronunciation information is determined from the phonetic alphabet. Is.

Therefore, it can be seen from this result that the recognition rate can be improved by applying the present invention to the words registered in the word dictionary information in advance.

In the above, in order to obtain pronunciation information, 2
I used the utterance once, but it may be one utterance, or
Two or more utterances may be used.

It is also possible to prepare two types, one with a noise signal superimposed and one without a noise signal, and register pronunciation information for each voice signal. The registration unit 11 has performed the registration process twice. That is, for one utterance, two types of voice signals, a voice signal passing through the voice synthesizing unit 33 and a voice signal not passing through the voice synthesizing unit 33, are registered.
Registration processing is performed in 1. Therefore, for example, in order to superpose and register a plurality of noises, the registration process of the registration unit 11 is performed a plurality of times.

In the above description, the noise in the running environment in the vehicle is used, but it is possible to deal with the noise in various environments such as office environment, airplane and train. Also,
In the case of voice recognition of a robot, it is possible to deal with a motor sound generated when the robot moves or a friction sound with a road surface generated when walking.

Furthermore, in the present embodiment, when the pronunciation information is registered, the registration unit 11 performs the registration process, and when the recognition is performed, the recognition unit 4 performs the recognition process. It is also possible to determine the pronunciation information simultaneously with the recognition processing. In this case, the word dictionary information of the language model database 6 is updated, if necessary.
That is, new words are added and pronunciation information is added to words already registered in the dictionary.

In the above description, the method of determining the pronunciation information from the input voice has been described, but it can be used in combination with other methods such as the registration of the pronunciation information by using another input means such as a keyboard.

The series of processes described above can be executed not only by hardware but also by software. In this case, for example, the voice recognition device 50 as shown in FIG. 10 is used.

In FIG. 10, the CPU (Central Proces
The sing unit) 51 is a program stored in a ROM (Read Only Memory) 52, or R from the storage unit 58.
Various processes are executed according to a program loaded in an AM (Random Access Memory) 53. The RAM 53 also appropriately stores data necessary for the CPU 51 to execute various processes.

CPU 51, ROM 52, and RAM 5
3 are connected to each other via a bus 54. An input / output interface 55 is also connected to the bus 54.

The input / output interface 55 includes an input unit 56 including a keyboard and a mouse and a CRT (Cathode R).
ay Tube), LCD (Liquid Crystal Display) and other displays, and output section 5 and speakers.
7, a storage unit 58 including a hard disk, a communication unit 5 including a modem, a terminal adapter, etc.
9 is connected. The communication unit 59 performs communication processing via a network (not shown).

A drive 60 is connected to the input / output interface 55 if necessary, and the magnetic disk 61,
The optical disc 62, the magneto-optical disc 63, the semiconductor memory 64, or the like is appropriately mounted, and the computer program read from the optical disc 62 is installed in the storage unit 58 as necessary.

When a series of processes are executed by software, a program that constitutes the software is installed in a computer in which dedicated hardware is installed, or various programs are installed to perform various functions. It is installed from a network or a recording medium into a general-purpose personal computer or the like capable of executing.

As shown in FIG. 10, this recording medium is a magnetic disk 61 (including a flexible disk) on which a program is recorded, which is distributed in order to provide the program to the user, separately from the apparatus main body, an optical disk. 62
(CD-ROM (Compact Disk-Read Only Memory), DVD (Dig
(including ital Versatile Disk)), magneto-optical disk 63
(Including MD (Mini-Disk) (trademark)), or package media consisting of semiconductor memory 64, etc. ROM
52 and a hard disk included in the storage unit 58.

In the present specification, the steps for describing the program recorded on the recording medium are not limited to the processing performed in time series according to the order described, but may be performed in parallel even if the processing is not necessarily performed in time series. Alternatively, it also includes processes that are individually executed.

[0088]

As described above, according to the voice recognition device and method, the recording medium, and the program of the present invention, the background noise is synthesized with the input voice, the synthesized voice is acoustically analyzed, and the feature amount of the synthesized voice is obtained. Is extracted, and the pronunciation information estimated based on the feature amount is registered as the pronunciation information of the corresponding word. Therefore, the recognition accuracy for background noise can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of a conventional voice recognition device.

FIG. 2 is a block diagram showing another configuration example of a conventional voice recognition device.

FIG. 3 is a diagram illustrating an acoustic model and pronunciation information of the voice recognition device in FIG.

FIG. 4 is a diagram illustrating an acoustic model network of the voice recognition device in FIG.

FIG. 5 is a block diagram showing a configuration example of a voice recognition device to which the present invention is applied.

6 is a flowchart illustrating a pronunciation information registration process of the voice recognition device in FIG.

7 is a flowchart illustrating a voice recognition process of the voice recognition device in FIG.

FIG. 8 is a diagram showing an experimental result of voice recognition using the voice recognition device of FIG.

9 is a diagram showing an example of pronunciation information of the voice recognition device in FIG.

FIG. 10 is a block diagram showing another configuration example of a voice recognition device to which the present invention has been applied.

[Explanation of symbols]

3 acoustic analysis unit, 4 recognition unit, 5 acoustic model database, 6 language model database, 11 registration unit, 1
2 acoustic model network, 33 noise storage unit, 34
Speech synthesizer

Claims (6)

[Claims] 1. A voice recognition device for performing voice recognition processing for recognizing an input voice, comprising: an acquisition unit that acquires background noise; and a synthesis that combines the input voice with the background noise acquired by the acquisition unit. Means, acoustically analyzing the synthesized voice synthesized by the synthesizing means,
An analysis unit that extracts a feature amount of the synthesized voice, an estimation unit that estimates pronunciation information based on the feature amount extracted by the analysis unit, and a pronunciation word that is estimated by the estimation unit And a registration means for registering the pronunciation information as the pronunciation information. 2. The voice recognition device according to claim 1, wherein the registration unit registers a plurality of the pronunciation information for the word. 3. The voice recognition device according to claim 1, further comprising a matching unit that performs matching processing based on the pronunciation information registered by the registration unit. 4. A voice recognition method for a voice recognition device that performs voice recognition processing for recognizing an input voice, comprising: an acquisition step of acquiring background noise; and the input voice acquired by the processing of the acquisition step. A synthesis step of synthesizing background noise, an analysis step of acoustically analyzing the synthesized speech synthesized by the processing of the synthesis step, and extracting a feature quantity of the synthesized speech, and the feature quantity extracted by the processing of the analysis step. A voice recognition method comprising: an estimation step of estimating pronunciation information based on the above; and a registration step of registering the pronunciation information estimated by the processing of the estimation step as the pronunciation information of a corresponding word. 5. A program for a voice recognition device that performs a voice recognition process of recognizing an input voice, the obtaining step of obtaining background noise, and the background obtained by the process of the obtaining step for the input voice. A synthesis step of synthesizing noise, an acoustic analysis of the synthesized speech synthesized by the processing of the synthesis step, an analysis step of extracting a feature amount of the synthesized speech, and the feature amount extracted by the processing of the analysis step A computer readable program comprising: an estimation step of estimating pronunciation information based on the pronunciation information; and a registration step of registering the pronunciation information estimated by the processing of the estimation step as the pronunciation information of a corresponding word. A recording medium on which a program is recorded. 6. A computer for a voice recognition device that performs a voice recognition process of recognizing an input voice, the acquisition step of acquiring background noise, and the background noise obtained by the process of the acquisition step of the input voice. A synthesizing step of synthesizing, an acoustic analysis of the synthesized speech synthesized by the processing of the synthesizing step, and an analysis step of extracting a feature amount of the synthesized voice, and based on the feature amount extracted by the processing of the analyzing step A program that executes an estimation step of estimating pronunciation information and a registration step of registering the pronunciation information estimated by the processing of the estimation step as the pronunciation information of a corresponding word.