JP4054610B2 - Voice recognition apparatus, voice recognition method, voice recognition program, and program recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition apparatus and speech recognition method using a subword as a recognition unit, a speech recognition program, and a program recording medium on which a speech recognition program is recorded.
[0002]
[Prior art]
Conventionally, in a speech recognition apparatus using a subword such as a phoneme or syllable as a recognition unit for an unspecified speaker, a method in which a user registers a recognition target word in a word dictionary using a character string, or a recognition target in advance In many cases, a method in which words are prepared in a word dictionary by character strings is adopted.
[0003]
The word dictionary used here is a subword notation dictionary in which a character string of each recognition target word is described by a network of subwords or a tree structure, and only the recognition target word artificially registered by the character string is stored in the word dictionary. Does not exist. Note that the speech recognition technology using the subword as a recognition unit is described in detail in the publication “Basics of Speech Recognition (2)” translated by Sadaaki Furui.
[0004]
However, voices uttered by humans are often different from registered character strings. For example, the longer sound of words such as “ei” and “ou”, that is, the longer sound of “e” and “o” is a common phenomenon. However, in a speech recognition apparatus that registers a recognition target word by a character string, a word dictionary corresponding to this longer sound is not created unless the user registers the recognition target word in consideration of the longer sound. is there.
[0005]
Therefore, there is a case where the recognition target word corresponding to the uttered voice is not registered, and the recognition rate decreases in the recognition target word that may be prolonged because there is an individual difference in particular regarding the lengthening of the sound. It is the cause that invites.
[0006]
In order to solve such problems, for recognition target words that contain syllables that may be prolonged when uttered, a speech that automatically registers a recognition target word that has been lengthened automatically from a character string in the word dictionary A recognition dictionary creation apparatus is disclosed in Japanese Patent Laid-Open No. 11-311991. According to this speech recognition dictionary creation device, a recognition target word of a character string that is automatically lengthened from a character string that may be uttered with a long sound is created and registered in the word dictionary. It is possible to prevent a recognition error due to. As a specific example, in the case of “Tokyo” including “ou”, which may have a longer sound, not only “Tokyo” but also “Tokyo”, “Tokyo”, “To- The three recognition target words “Kyo” are also automatically registered in the word dictionary.
[0007]
[Problems to be solved by the invention]
However, the speech recognition device equipped with the speech recognition dictionary creation device disclosed in the above-mentioned JP-A-11-311991 has the following problems. That is, in the speech recognition dictionary creating apparatus, as described above, when there are n places that may be made longer in one word, 2 nth recognition target words are registered. Will be. Therefore, if the recognition target word corresponding to the longer sound is automatically added and registered in the word dictionary by such dictionary creation processing, a new problem arises that as a result, the number of recognition target words increases and it takes time to collate the dictionary. Will do.
[0008]
SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to provide a speech recognition apparatus and speech recognition method, speech recognition program, and program recording medium on which a speech recognition program is recorded, which can cope with a longer sound without increasing the recognition target word and processing time. There is.
[0009]
[Means for Solving the Problems]
To achieve the above object, the speech recognition apparatus of this invention, the acoustic model storage unit subword acoustic model is stored an acoustic model of Sa Buwado, recognition target represented by sub-word sequence is a time series of sub Buwado It has a word dictionary in which words are stored , and the above subword has a special notation that is a combination of a notation that is not a long sound and a notation that is a long sound. Contains the subword represented . Then, the collation unit collates the time series of the characteristic parameters obtained from the input speech with the subword acoustic model, and obtains the likelihood of each subword. Further, the likelihood of each subword obtained is compared with the subword sequence of the recognition target word, and the recognition target word having the highest likelihood is generated as a recognition result.
[0010]
Thus, sub-word acoustic models of sub-word containing the word represented in a special notation with represents the possibility of a prolonged sound of the recognition target words represented by sub-word sequences comprising a word represented by a special notation Therefore, the sub-word acoustic model and the recognition target word can be used both when the syllable that may be prolonged is long and when it is not long. Therefore, there is a word in the input voice that may be prolonged, and it can be easily dealt with whether the word is prolonged or not, and a recognition error due to the prolonged sound. Is prevented.
[0011]
At that time, a syllable that may be lengthened in the recognition target word stored in the word dictionary is represented by one special notation subword. Therefore, a subword that represents a syllable that is lengthened as in the past. And a subword representing a syllable that is not lengthened, the number of subword sequences representing the recognition target word is reduced. Therefore, the processing time for matching the likelihood of each subword with the subword sequence of the recognition target word is shortened compared to the conventional case.
[0012]
In the speech recognition apparatus of one embodiment, the dictionary creation unit detects a syllable that may be converted into a subword notation by converting the character string notation of the input recognition target word into a subword notation. Further, the detected syllables having the possibility of lengthening are converted into subwords represented by the special notation, and a subword sequence representing the input recognition target word is obtained. Then, the recognition target words represented by the obtained subword series are stored in the word dictionary.
[0013]
In this way, one recognition target word stored in the word dictionary is represented by one subword sequence, and the processing time when matching the likelihood of each subword with the subword sequence of the recognition target word is minimized. It is done.
[0014]
Further, in the speech recognition apparatus of one embodiment, the dictionary creating means converts the character string notation of the input recognition target word mixed with kanji characters into kana character notation and converts the obtained kana character notation into subword notation. I have to. Therefore, it is not necessary for the user to input the recognition target word by reading, and it is possible to easily register even when the recognition target word of the large vocabulary is registered.
[0015]
In the speech recognition apparatus of one embodiment, the dictionary creating means converts the kana character notation of the input recognition target word into a subword notation. Therefore, there is no need to mount the kanji kana conversion means in the dictionary creation means, and a compact speech recognition device is realized.
[0016]
The voice recognition method of the present invention, the time series of feature parameters obtained from the input speech, a word of the possibility of a prolonged sound reduction, notation conventions and prolonged sound reduction has been case of If not prolonged sound of Are compared with a subword acoustic model of a subword including a subword represented by a special notation that is written together , and the likelihood of each subword is obtained. Further, the likelihood of each subword obtained is compared with the recognition target word represented by the subword sequence of the subword including the subword represented by the special notation, and the recognition target exhibiting the highest likelihood A word is generated as a recognition result.
[0017]
Thus, at the time of the above collation, a subword acoustic model and a recognition target word that are compatible with both cases where a syllable that may have a longer sound length becomes longer and when it is not longer sounded are used. Therefore, there is a word in the input voice that may be prolonged, and it can be easily dealt with whether the word is prolonged or not, and a recognition error due to the prolonged sound. Is prevented.
[0018]
At that time, a syllable having a possibility of lengthening in the recognition target word stored in the word dictionary is represented by one special notation subword. Therefore, the processing time for matching the likelihood of each subword with the subword sequence of the recognition target word is shortened compared to the conventional case where no subword representing the possibility of longer sound is used.
[0019]
The speech recognition program of the present invention can cause a computer to function as acoustic analysis processing means, acoustic model storage means, word dictionary, and collation means in the speech recognition apparatus of the present invention. Therefore, it is possible to easily cope with a case where a word having a possibility of a longer sound existing in the input speech is made longer or not, and a recognition error due to the longer sound is prevented. Furthermore, the processing time for matching the likelihood of each subword with the subword sequence of the recognition target word is shortened compared to the conventional case.
[0020]
The program recording medium of the present invention records the voice recognition program of the present invention. Therefore, by reading out and executing by a computer, a recognition error is prevented when a word having a possibility of lengthening existing in the input voice is lengthened. Furthermore, the processing time for matching the likelihood of each subword with the subword sequence of the recognition target word is shortened compared to the conventional case.
[0021]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments. FIG. 1 is a block diagram of the speech recognition apparatus according to the present embodiment. This speech recognition apparatus includes an acoustic analysis unit 1, a likelihood calculation unit 2, an acoustic model storage unit 3, a likelihood storage unit 4, a dictionary collation unit 5, a word dictionary 6, a dictionary creation unit 7, and a kana character / phoneme correspondence table 8. Consists of.
[0022]
In FIG. 1, the input speech is converted into a characteristic parameter series by the acoustic analysis unit 1 and output to the likelihood calculation unit 2. In the likelihood calculation unit 2, the acoustic model of each phoneme stored in the acoustic model storage unit 3 is collated with the input feature parameter series, and the calculated likelihood of each phoneme is stored in the likelihood storage unit 4. Stored. The dictionary collation unit 5 collates the phoneme sequence of each word in the word dictionary 6 with the likelihood of each phoneme stored in the likelihood storage unit 4 using the Viterbi algorithm, and each word Is calculated. Then, the word with the highest likelihood is output as the recognition result. That is, in the present embodiment, the likelihood calculation unit 2 and the dictionary collation unit 5 constitute the collation unit.
[0023]
As the acoustic model, for example, a hidden Markov model (HMM) called a monophone model that does not depend on the surrounding phonemic environment is used. However, in the present embodiment, two phonemes of phoneme “i / e” and phoneme “u / o” are added to the above-mentioned HMM phonemes as special phonemes indicating the possibility of a longer tone. .
[0024]
In the likelihood storage unit 4, as described above, matching between the acoustic model of each phoneme stored in the acoustic model storage unit 3 and the feature parameter series input from the acoustic analysis unit 1 is performed by the likelihood calculation unit 2. When this is done, the calculated likelihood of each phoneme is sequentially stored.
[0025]
The word dictionary 6 stores a phoneme sequence of each word created by the dictionary creation unit 7 based on the recognition target word in the form of a phoneme network or a phoneme tree structure. The phoneme sequence of this word is a phoneme sequence of the reading of each recognition target word. In that case, as shown in the rightmost column of FIG. 2 (a), a special phoneme “i / e” indicating the possibility of lengthening is used for phonemes that may be lengthened in the phoneme series. ”,“ U / o ”.
[0026]
Here, the phoneme sequence “t; o; u; k; j; o; u” in FIG. 2 (a) is replaced with the conventional notation method using the special phonemes “i / e” and “u / o”. If phonemes are expressed in a format that can support longer sound, the four phonemes are displayed as shown in FIG. This is the same as FIG. 2 (a) as shown in FIG. 2 (b) when the phoneme is expressed by the notation method of the present embodiment using the special phonemes “i / e” and “u / o”. One phoneme notation is sufficient, and the number of phoneme notations by the conventional notation method can be reduced. Accordingly, the collation calculation processing by the dictionary collation unit 4 using the phoneme sequence of each word stored in the word dictionary 6 can be reduced, and the collation time can be shortened.
[0027]
Hereinafter, a method for creating a phoneme sequence of a word registered in the word dictionary 6 performed by the dictionary creation unit 7 will be described in detail according to a flowchart of the dictionary creation processing operation shown in FIG. When a recognition target word is input to the dictionary creation unit 7, the dictionary creation processing operation starts.
[0028]
In step S1, first, a recognition target word written with characters mixed with kanji characters is converted into a kana character string using a kanji kana conversion technique. For example, the input recognition target word “Tokyo” is converted into a kana character string “Tokyo”. In step S2, the obtained kana character string is converted into a phoneme string with reference to a kana character / phoneme correspondence table 8 as shown in FIG. As a result, the kana character string “Tokyo” is converted into a phoneme string “t; o; u; k; j; o; u”.
[0029]
In step S3, phoneme chains “e; i”, “e; e”, “o; u”, and “o; o” are detected from the phoneme sequence obtained in step S2. In this way, syllables that may be prolonged are extracted. As a result, two phoneme chains “o; u” are detected from the phoneme sequence “t; o; u; k; j; o; u”, and the syllable “u” is extracted as a syllable that may be prolonged. Is done.
[0030]
In step S4, the syllable with the possibility of lengthening extracted in step S3 is replaced with the special phoneme notations “i / e” and “u / o” representing the possibility of lengthening. That is, the phoneme chains “e; i” and “e; e” are replaced with “e; i / e”, while the phoneme chains “o; u” and “o; o” are replaced with “o; u / o”. ". As a result, in the case of the phoneme string “t; o; u; k; j; o; u”, the phoneme string “t; o; u / o; k; j; o; u / o” is obtained. In step S5, the phoneme string obtained in step S4 is converted into data corresponding to the collation method in the dictionary collation unit 5. After that, the dictionary creation processing operation is terminated.
[0031]
FIG. 5 is an example in which each word registered in the word dictionary 6 is expressed in the form of a phoneme network. 5A shows a phoneme network created by the dictionary creation unit 7 based on the input recognition target words “Hokkaido”, “Aomori”,..., “Tokyo”,. FIG. 5B shows a phoneme network created based on the input recognition target words “Hokkaido”,..., “Tokyo” by a conventional method.
[0032]
FIG. 6 is an example in which each word registered in the word dictionary 6 is expressed in a phoneme tree structure format. In FIG. 6 (a), the dictionary creation unit 7 uses the input recognition target words "Akita", "Aomori", ..., "Hokkaido", ..., "Oita", "Osaka", ..., "Tokyo", ... Is a phoneme tree structure created based on FIG. 6B shows a phoneme tree structure created based on the input recognition target words “Hokkaido”,..., “Tokyo” by the conventional method.
[0033]
Therefore, when the dictionary collation unit 5 performs collation using the word dictionary 6 created by the dictionary creation unit 7, the phoneme network or the tree structure search path constituting the dictionary is minimized. As a result, the processing time required for dictionary collation can be significantly shortened compared to conventional word dictionaries.
[0034]
As described above, in the present embodiment, in addition to the conventional phoneme HMM, the acoustic model storage unit 3 includes two special phonemes “i / e” and “u / o ”HMM is added. Also, in the word dictionary 6, the dictionary creation unit 7 describes the syllables that may exist in each recognition target word as the special phonemes “i / e” and “u / o”. The word dictionary created by
[0035]
In this way, the syllables with the possibility of lengthening in the recognition target words registered in the word dictionary 6 are represented by special phonemes that can be used both when the sound is made longer and when it is not made longer. Yes. Therefore, there is a word that may become longer sound in the input speech, and the recognition rate does not decrease even if the word is made longer or not longer.
[0036]
Further, for example, when 47 prefecture names are used as recognition target vocabulary, and when dealing with lengthening by a conventional recognition method, each phoneme when the same recognition target word is expressed by a plurality of phoneme strings is used. When there are n places where there is a possibility of making the sound longer in one recognition target word, such as “Hokkaido” and 2 “Tokyo” as shown in FIG. 2 (b). Needs to register 2 n words in the word dictionary. Therefore, the number of words that are considerably larger than the number of prefectures “47” is registered in the word dictionary, which greatly increases the processing time for dictionary collation. On the other hand, in the case of the present embodiment, since one prefecture name can be expressed by one word, the number of words registered in the word dictionary 6 is “47”, which is the same as the number of prefecture names. . Therefore, the dictionary collation processing time hardly increases with respect to the minimum necessary processing time.
[0037]
In the above embodiment, when the dictionary creation unit 7 executes the dictionary creation processing operation, the recognition target word represented by the kanji-kana mixed characters is used in step S1 using the kanji kana conversion technique. After the conversion to the kana character string, the phoneme notation string is created in step S2. However, the present invention is not limited to this. For example, when the reading of the word to be recognized is input to the dictionary creation unit 7 using kana characters, the step S1 can be omitted and the dictionary creation processing operation can be started from the step S2.
[0038]
In the dictionary creation process by the dictionary creation unit 7 in the above embodiment, conversion from a kana character string to a phoneme notation string directly is performed. However, the kana character string may be converted once from the romaji character string, and the obtained romaji character string may be converted into the phoneme notation string.
[0039]
In addition, the dictionary creation process by the dictionary creation unit 7 in the above embodiment may be separately performed by a computer using a dictionary creation program that can cause the computer to execute a dictionary creation processing operation similar to the dictionary creation unit 7. There is no problem.
[0040]
The acoustic model stored in the acoustic model storage unit 3 in the above embodiment is an HMM that does not depend on the phoneme environment before and after that is called a monophone model. However, the acoustic model is not limited to this, and the subword of the acoustic model may be a syllable or an HMM that depends on the surrounding phonemic environment.
[0041]
By the way, the acoustic analysis processing means, acoustic model storage means, word dictionary and collation means by the acoustic analysis section 1, likelihood calculation section 2, acoustic model storage section 3, dictionary collation section 5 and word dictionary 6 in the above embodiments. Is realized by a voice recognition program recorded on a program recording medium. The program recording medium in each of the above embodiments is a program medium composed of a ROM (read only memory). Alternatively, it may be a program medium that is loaded into an external auxiliary storage device and read out. In any case, the program reading means for reading the voice recognition program from the program medium may have a configuration in which the program medium is directly accessed and read, or in a RAM (Random Access Memory). You may have the structure which downloads to the program storage area (not shown) provided, and accesses and reads the said program storage area. It is assumed that a download program for downloading from the program medium to the program storage area of the RAM is stored in the main unit in advance.
[0042]
Here, the program medium is configured to be separable from the main body side, and is a tape system such as a magnetic tape or a cassette tape, a magnetic disk such as a floppy disk or a hard disk, a CD (compact disk) -ROM, or MO (magneto-optical). Optical discs such as discs, MD (mini discs) and DVDs (digital versatile discs), card systems such as IC (integrated circuit) cards and optical cards, mask ROM, EPROM (ultraviolet erasable ROM), EEPROM (electrical This is a medium that carries a fixed program, including a semiconductor memory system such as a static erasable ROM) and a flash ROM.
[0043]
In addition, the speech recognition apparatus in each of the above embodiments can be configured to include a modem and be connectable to a communication network including the Internet. In this case, the program medium may be a medium that dynamically carries the program by downloading from a communication network or the like. In this case, it is assumed that a download program for downloading from the communication network is stored in the main device in advance. Or it shall be installed from another recording medium.
[0044]
It should be noted that what is recorded on the recording medium is not limited to a program, and data can also be recorded.
[0045]
【The invention's effect】
As is clear from the above, in the present invention, the time series of the characteristic parameters obtained from the input speech and the subwords with the possibility of longer sound, the notation when not being made longer and the notation when being made longer performing matching between the sub-word acoustic models of sub-word containing the word, represented by special notation comprising also shown the door, give the likelihood of each word, and the likelihood of each word obtained, the special Since the recognition target word having the highest likelihood is generated as a recognition result by performing matching with the recognition target word represented by the subword series of subwords including the subword represented by the notation , In the input voice, there is a word that has a possibility of making a long sound, and it is possible to easily cope with a case where the word is made long or not. Therefore, it is possible to prevent a recognition error when the syllable in the voice input word is lengthened.
[0046]
At that time, syllables having a possibility of lengthening in the recognition target words stored in the word dictionary are described by one special notation subword per syllable. Therefore, one recognition target word can be represented by only one subword sequence, and the number of subword sequences and the processing time required for matching the likelihood of each subword with the subword sequence of the recognition target word are increased. Can be suppressed as much as possible. As a result, the above collation processing time can be shortened as compared with the conventional speech recognition in which one recognition target word is represented by a plurality of subword sequences to cope with the longer sound.
[Brief description of the drawings]
FIG. 1 is a block diagram of a speech recognition apparatus according to the present invention.
2 is an explanatory diagram of a phoneme sequence of words stored in the word dictionary in FIG. 1. FIG.
FIG. 3 is a flowchart of a dictionary creation processing operation executed by the dictionary creation unit in FIG. 1;
4 is a diagram showing an example of a concept of a kana character / phoneme correspondence table in FIG. 1; FIG.
5 is a diagram illustrating an example in which each word registered in the word dictionary in FIG. 1 is expressed in the form of a phoneme network.
6 is a diagram illustrating an example in which each word registered in the word dictionary in FIG. 1 is expressed in a phoneme tree structure format. FIG.
[Explanation of symbols]
1 ... acoustic analysis section,
2 ... Likelihood calculation part,
3 ... Acoustic model storage unit,
4 ... Likelihood storage unit,
5 ... Dictionary collation part,
6 ... Word dictionary,
7 ... Dictionary creation part,
8: Kana character / phoneme correspondence table.

Claims (7)

A speech recognition device that recognizes an input speech using a subword as a recognition unit and a subword acoustic model,
Acoustic analysis processing means for analyzing the input speech and obtaining a time series of feature parameters ;
An acoustic model storage means for the sub-word acoustic models are stored an acoustic model of service Buwado,
A word dictionary recognized word is stored represented in a sub-word sequence is a time series of sub Buwado,
The likelihood of each subword is obtained by collating the time series of the characteristic parameters with the subword acoustic model, and the likelihood of each subword is collated with the subword sequence of the recognition target word to obtain the most likely It has collation means to generate a high recognition target word as a recognition result
The subword includes a subword that has a possibility of making a long sound and is expressed by a special notation in which a notation when not making a long sound and a notation when making a long sound are written together. <br/> A speech recognition apparatus characterized by the above. The speech recognition apparatus according to claim 1,
A recognition target word written in a character string is input, the character string representation of this input recognition target word is converted into a subword representation, and a syllable that may be made longer is detected from the obtained subword representation, and this detection is performed. A dictionary creating means for converting a syllable that has a possibility of lengthening into a subword represented by the special notation and storing a recognition target word represented by the obtained subword sequence in the word dictionary A voice recognition device characterized by the above. The speech recognition device according to claim 2,
The dictionary creation means inputs a recognition target word written in a kanji-kana mixed character string, converts the kanji kana-mixed character string representation of the input recognition target word into a kana character notation, and obtains the obtained kana character notation A speech recognition apparatus characterized in that it is converted to subword notation. The speech recognition device according to claim 2,
A speech recognition apparatus, wherein the dictionary creating means receives a recognition target word expressed in a kana character string, and converts the kana character expression of the input recognition target word into a subword expression. A speech recognition method using a subword as a recognition unit and recognizing input speech using a subword acoustic model,
Analyzing the input speech to obtain a time series of feature parameters;
A word of the possibility of a prolonged sound of the sub-word, an acoustic model of the sub-word containing the word represented in a special notation comprising shown together with notation when labeled with long sound of If not prolonged sound of and acoustic models by performing a collation with the time series of the feature parameters, obtaining a likelihood of each sub-word,
Recognize the word with the highest likelihood by collating the recognition target word represented by the subword sequence that is the time series of the subwords including the subword represented by the special notation with the likelihood of each subword. A speech recognition method comprising the step of generating as a result. Computer
A speech recognition program that functions as an acoustic analysis processing means, an acoustic model storage means, a word dictionary, and a collation means according to claim 1.   A computer-readable program recording medium on which the voice recognition program according to claim 6 is recorded.

Images