JP2004138914A - Device and method for adjusting dictionary for voice recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device and a method for adjusting a dictionary for voice recognition capable of efficiently performing maintenance of the dictionary for the voice recognition while improving the recognition accuracy. <P>SOLUTION: The device and the method input voice signals generated by a user, recognize them by using a first dictionary for voice recognition structured by each word of blocks and extracts at least the voice signals of each word of the blocks to be recognized and generating times of the voice signals from the recognition result. The device and the method recognize voices again by using a second dictionary for voice recognition structured by each word by using corresponding voice signals at each extracted generation time, and based on the recognition result, stores reading information corresponding to the voice signals of each word of the blocks as a pair of data and updates the first dictionary by using the pair of the data. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech recognition dictionary adjustment device and method capable of efficiently maintaining a speech recognition dictionary in a speech recognition device.
[0002]
[Prior art]
In the conventional speech recognition device, a word for the highest evaluation value and reading information corresponding to the word with the highest evaluation value are aggregated as a pair of data for the purpose of improving recognition accuracy, thereby forming a speech recognition dictionary for the input speech signal. are doing.
[0003]
However, in the conventional speech recognition device, the reading information of the word to be recognized to be registered in the speech recognition dictionary is often manually performed by an expert, but such work is complicated in man-hours. Therefore, a technique of adding reading information in advance by defining a specific automatic conversion rule or the like is widely used.
[0004]
However, there are fluctuations in reading individual words, even for the same speaker. For example, even with the same speaker, the same word, that is, a word described as “Honorific” may be uttered as “Keigo” or “Kego”. Therefore, even in such a case, in order to accurately perform speech recognition, the speech recognition accuracy is improved by manually adding a reading to each word or changing a conversion rule by an expert. Tuning work was indispensable.
[0005]
However, it is impractical to perform such a task manually, even for an expert. Therefore, many measures have been considered for automatically performing tuning by adopting a specific system.
[0006]
For example, in (Patent Document 1), a voice recognition mode and a voice registration mode are switched. In the voice registration mode, a voice signal to be registered in a word unit is also divided into constituent word units and stored in a voice recognition dictionary. A technique has been disclosed in which a large number of registration data can be obtained by one registration operation by registering.
[0007]
Also, in Patent Document 2, by automatically assigning a plurality of readings to one character (word), it is possible to reliably estimate recognized characters even for readings that have not been registered. There is disclosed a technology that can do this.
[0008]
[Patent Document 1]
JP-A-11-282486
[Patent Document 2]
JP 2000-47684 A
[Problems to be solved by the invention]
However, in the method disclosed in (Patent Document 1), there is a problem in accuracy of expanding words in word units, and in some cases, unregistered syllables or phonemes may be generated. However, there is a problem that the recognition accuracy is lowered.
[0011]
In addition, according to the method disclosed in Patent Document 2, it is difficult to determine to what extent a plurality of readings are added to improve recognition accuracy, and after all, it is difficult for an expert to maintain a speech recognition dictionary. However, there is a problem that the number of working steps is equivalent to that of the above.
[0012]
SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech recognition dictionary adjustment apparatus and method capable of efficiently performing maintenance of a speech recognition dictionary while improving recognition accuracy in order to solve the above problems.
[0013]
[Means for Solving the Problems]
In order to achieve the above object, a voice recognition dictionary adjustment device according to the present invention includes a voice signal input unit for inputting a voice signal emitted by a user, and a voice signal input unit configured to convert the input voice signal into a group of words. A first speech recognition unit for recognizing by using the first speech recognition dictionary, and at least a set of speech signals in units of words and a generation time of the speech signal, which are recognition targets, based on the recognition result in the first speech recognition unit. An audio signal information extracting unit to be extracted, a second audio recognizing unit that recognizes again by using a second audio recognition dictionary configured for each word using a corresponding audio signal for each occurrence time, A recognition result storage unit that stores a set of speech signals corresponding to a group of words and corresponding reading information as a pair of data, based on the recognition result of the second voice recognition unit; Characterized in that it comprises a recognition dictionary update unit for updating the speech recognition dictionary.
[0014]
With this configuration, speech recognition processing is performed in units of words to extract speech data in units of words, and then speech recognition corresponding to the words is performed again in units of words, so that unregistered syllables and phonemes are obtained. Since the words used by the user can be reliably added to the dictionary for speech recognition without any occurrence of the like, an efficient dictionary for speech recognition with high speech recognition accuracy without registering useless reading information. It can be adjusted so that
[0015]
In addition, it is preferable that the speech recognition dictionary adjustment device according to the present invention include an update number counting unit that counts the number of times the first speech recognition dictionary has been updated for each group of words. This is because a recognition result having a small number of update times is considered to be likely to be erroneous recognition, so that it is possible to prevent the recognition result from being reflected in the speech recognition dictionary.
[0016]
Further, the speech recognition dictionary adjustment device according to the present invention preferably stores the first speech recognition dictionary for each user. Alternatively, the speech recognition dictionary adjustment device according to the present invention preferably stores the first speech recognition dictionary for each environment used by the user.
[0017]
It is also preferable that the speech recognition dictionary adjustment device according to the present invention uses a plurality of second speech recognition engines.
[0018]
Further, the present invention is characterized by software that executes the function of the above-described dictionary adjustment device for speech recognition as a processing step of a computer, and more specifically, a software for inputting a speech signal emitted by a user. One step, a second step of recognizing the input voice signal using a first voice recognition dictionary configured in a unit of words, and at least a recognition target based on a recognition result in the second step. A third step of extracting a voice signal in units of words and a generation time of the voice signal, and a second voice recognition unit configured for each generation time using a corresponding voice signal for each word. A fourth step of re-recognizing using the dictionary, and a fifth step of storing a set of speech signals in word units and corresponding reading information as a pair of data based on the recognition result in the fourth step. And a sixth step of updating the first dictionary for speech recognition using a pair of stored data, and a method for adjusting the dictionary for speech recognition, and a computer-executable program embodying such a step. There is a feature.
[0019]
With this configuration, by loading and executing the program on a computer, speech recognition processing is performed in units of words to extract speech data in units of words, and thereafter, speech recognition in units of words is performed on speech data corresponding to the words. Again, no unregistered syllables, phonemes, etc. are generated, and the words used by the user can be reliably added to the speech recognition dictionary. In addition, it is possible to realize a speech recognition dictionary adjustment device that can be adjusted to be an efficient speech recognition dictionary with high speech recognition accuracy.
[0020]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, a dictionary adjustment device for speech recognition according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram of a speech recognition dictionary adjustment device according to an embodiment of the present invention.
[0021]
In FIG. 1, when a user 11 who inputs voice inputs the word “today's weather” by voice, first, the first voice recognition engine 12 outputs a first voice composed of a group of words. Voice recognition is performed with reference to the recognition dictionary 13.
[0022]
FIG. 2 shows an example of a data configuration of the first speech recognition dictionary 13 in the speech recognition dictionary adjustment device according to the embodiment of the present invention. In the example of FIG. 2, the first speech recognition dictionary 13 is configured as a set of data indicating the correspondence with the reading information for each part of speech of a word. Of course, it is not limited to a part of speech unit.
[0023]
Next, the recognition result extraction unit 14 extracts a word output as a recognition result and a speech signal corresponding to the output word. The extraction of the audio signal is performed by a temporal signal section corresponding to the output word. For example, in the example of FIG. 1, the signal section corresponding to “today” is a section between 10 ms and 200 ms, the signal section corresponding to “no” is a section between 200 ms and 250 ms, and the signal section corresponding to “weather” is A section between 250 ms and 500 ms is extracted.
[0024]
Then, the second speech recognition engine 15 refers to the second speech recognition dictionary 16 composed of one word such as syllables and phonemes, and performs speech recognition for each speech data section in units of reading information. I do. For example, assuming that speech is recognized in units of syllables, in the example of FIG. 1, two syllables “Kyo” and “O” are based on a voice signal in a section between 10 ms and 200 ms corresponding to “Today”. Is recognized by the second speech recognition engine 15.
[0025]
FIG. 3 shows an example of the data configuration of the second speech recognition dictionary 16 in the speech recognition dictionary adjustment device according to the embodiment of the present invention. In the example of FIG. 3, the second speech recognition dictionary 16 is configured as a set of a pair of data indicating a correspondence with reading information in each syllable unit such as “A”, “I”, and “U”. Of course, the present invention is not limited to syllable units, but may be, for example, phoneme units.
[0026]
Then, the recognition result storage unit 17 stores the recognition result by the second voice recognition engine 15 as a pair of data associated with a word corresponding to a voice signal corresponding to a signal section to be recognized. That is, as shown in FIG. 4, "Kyoo" is stored in the recognition result storage unit 17 as reading information for the word "today". Similarly, for all input speech signals, the recognition result by the second speech recognition engine 15 is used as a pair of data associated with a word corresponding to the speech signal corresponding to the signal section to be recognized. Will be saved.
[0027]
Finally, the recognition dictionary updating unit 18 updates the content of the speech recognition dictionary with respect to the first speech recognition dictionary 13 based on the pair of data stored in the recognition result storage unit 17. .
[0028]
Here, as shown in FIG. 5, a plurality of speech recognition engines may be used as the second speech recognition engine 15. In the example of FIG. 5, a third speech recognition engine 19 is further added to the second speech recognition engine 15.
[0029]
The second speech recognition engine 15 and the third speech recognition engine 19 use different acoustic models. Then, while referring to the second speech recognition dictionary 16 composed of single words such as syllables and phonemes, speech recognition is performed for each speech data section in units of reading information.
[0030]
For example, in the example of FIG. 5, for the voice data section associated with the word “temperature”, the second voice recognition engine 15 is “Kyog” and the third voice recognition engine 19 is “Tenki”. When the recognition is performed with different reading information, since the recognition results are different in a plurality of speech recognition engines, it is determined that the recognition result is incorrect, and the first speech recognition dictionary 13 is updated. That is not to do.
[0031]
On the other hand, when the voice data section associated with the word “today” is recognized as “Kyo” by the second voice recognition engine 15 and “Kyo” by the third voice recognition engine 19, Since the recognition results match in a plurality of speech recognition engines, the recognition result is determined to be correct, and the first speech recognition dictionary 13 is updated.
[0032]
When the reading information portion of the recognition result of “Temperature: Tenki” by the third voice recognition engine 19 is checked against the first voice recognition dictionary 13, the reading information portion is already registered. It is conceivable that it matches the pair of data "weather: weather". In such a case, the first voice recognition dictionary 13 may not be updated.
[0033]
In other words, when the first dictionary 13 for speech recognition is updated by the recognition dictionary update unit 18, it is checked whether or not there is any overlap with reading information corresponding to other words existing in the first dictionary 13 for speech recognition. become.
[0034]
FIG. 6A shows an example of the data configuration of the first dictionary for speech recognition 13 in the case of performing such matching, and FIG. 6B shows an example of the data configuration of the recognition result storage unit 17, respectively.
[0035]
As shown in FIG. 6B, for example, “Jiro”, which is a reading information portion of a pair of data “Taro: Jiro” in the recognition result storage unit 17, is already registered in the first voice recognition dictionary 13. It can be detected that it also exists in another pair of data “Jiro: Jiro”. In this case, the pair of data “Taro: Jiro” may be deleted, or “Taro” and “Jiro” may be presented as words having similar reading information.
[0036]
In addition, when the same reading information exists in the first dictionary 13 for speech recognition, for example, as shown in FIG. 6B, “Yamada: ga still” and “Kamata: ga still” are stored in the recognition result storage 17. If two pairs of data exist, both data may be deleted. Alternatively, it is conceivable to determine whether to delete words by verifying whether or not words having the same reading information appear simultaneously in the grammar.
[0037]
Further, when the first dictionary 13 for speech recognition is updated by the recognition dictionary update unit 18, the same pair of data as the pair of data to be updated is included in a pair of data of reading information corresponding to a word already stored. It is also conceivable that the data exists. In this case, the frequency at which the pair of data is stored is totaled, and a pair of data of the word to be added to the first dictionary 13 for speech recognition and the reading information of the word is determined based on the totalized update frequency. It is also possible to do.
[0038]
FIG. 7A shows an example of a data configuration of the first dictionary 13 for speech recognition in the case of performing such matching, and FIG. 7B shows an example of a data configuration of the recognition result storage unit 17, respectively.
[0039]
For example, the pair of data stored in the recognition result storage unit 17 includes six pairs of data whose update frequency is 100 times or more. Of these, the first voice recognition dictionary 13 includes Registering, in the first voice recognition dictionary 13, three pairs of data that are not included, “Today: ky o o”, “Tomorrow: a su ta”, and “Weather: teg ki” become. As described above, only the frequently updated data is set as the target of updating the first voice recognition dictionary 13, so that the result of accidental recognition by mistake is not reflected on the first voice recognition dictionary 13. Therefore, the dictionary can be updated without lowering the speech recognition accuracy of the updated first speech recognition dictionary 13.
[0040]
It is also conceivable that the first voice recognition dictionary 13 is stored for each user. That is, as shown in FIG. 8, a first voice recognition dictionary 13 is formed for each user ID, and the first voice recognition dictionary 13 is recognized by a user ID recognition unit 20 that selects a button in which the user ID is described, for example. The first dictionary for voice recognition 13 is updated for each user ID that has been set.
[0041]
Further, the present invention is not particularly limited to holding for each user. For example, the present invention relates to an environment for inputting audio signals such as a background noise level in a user's use environment, a type of telephone line used, and a type of microphone. A plurality of first speech recognition dictionaries 13 may be provided based on the information.
[0042]
Next, a description will be given of a processing flow of a program for realizing the speech recognition dictionary adjustment device according to the embodiment of the present invention. FIG. 9 shows a flowchart of the processing of a program for realizing the speech recognition dictionary adjustment device according to the embodiment of the present invention.
[0043]
In FIG. 9, first, a user's voice signal is received (step S901), and the first voice recognition engine 12 refers to the first voice recognition dictionary 13 composed of a group of words to generate a voice. Recognition is performed (step S902).
[0044]
Next, a word output as a recognition result and a speech signal corresponding to the output word are extracted as a signal section (step S903). Then, the second speech recognition engine 15 refers to the second speech recognition dictionary 16 composed of one word such as syllables and phonemes, and performs speech recognition for each speech data section in units of reading information. Is performed (step S904).
[0045]
Then, the recognition result by the second speech recognition engine 15 is stored in association with the word corresponding to the speech data to be recognized (step S905), and the contents of the first speech recognition dictionary 13 are stored. The content is updated based on the content (step S906).
[0046]
In the present embodiment, the first speech recognition engine 12 and the second speech recognition engine 15 or the other speech recognition engines to be added may use the same speech recognition engine, or may use different speech recognition engines. A recognition engine may be used. Further, in the present embodiment, the first speech recognition dictionary 13 is updated based only on the reading information determined to be the best, but the first speech recognition dictionary 13 is updated for a plurality of reading information candidates. The dictionary 13 may be updated.
[0047]
As described above, according to the present embodiment, speech data is extracted in units of words by performing speech recognition processing in units of words, and speech recognition is performed again in units of words for speech data corresponding to words. Since unregistered syllables and phonemes do not occur, the words used by the user can be reliably added to the dictionary for speech recognition. It is possible to make adjustments so as to be a highly efficient dictionary for speech recognition.
[0048]
As shown in FIG. 10, the program for realizing the speech recognition dictionary adjustment device according to the embodiment of the present invention includes not only the portable recording medium 102 such as the CD-ROM 102-1 and the flexible disk 102-2 but also the communication The program may be stored in any of the other storage device 101 provided at the end of the line and the recording medium 104 such as the hard disk or RAM of the computer 103. When the program is executed, the program is loaded and stored in the main memory. Executed in
[0049]
Also, as shown in FIG. 10, the first speech recognition dictionary and the second speech recognition dictionary used by the speech recognition dictionary adjusting device according to the embodiment of the present invention also have the CD-ROM 102-1. And any other storage device 101 provided at the end of the communication line, or a storage medium 104 such as a hard disk or a RAM of the computer 103, as well as the portable storage medium 102 such as the hard disk and the flexible disk 102-2. For example, it is read by the computer 103 when the speech recognition dictionary adjustment device according to the present invention is used.
[0050]
(Supplementary Note 1) An audio signal input unit for inputting an audio signal emitted by the user;
A first voice recognition unit that recognizes the input voice signal using a first voice recognition dictionary configured in a unit of words;
An audio signal information extraction unit that extracts an audio signal in units of words and a generation time of the audio signal, which are at least recognition targets, from a recognition result in the first audio recognition unit;
For each occurrence time, a second speech recognition unit that recognizes again using a second speech recognition dictionary configured for each word using the corresponding speech signal,
A recognition result storage unit for storing the speech signal in units of words and corresponding reading information as a pair of data based on a recognition result in the second speech recognition unit. Dictionary adjustment device.
[0051]
(Supplementary Note 2) The speech recognition dictionary adjustment device according to supplementary note 1, further comprising a recognition dictionary updating unit that updates the first speech recognition dictionary using the stored pair of data.
[0052]
(Supplementary Note 3) The speech recognition dictionary adjustment device according to Supplementary Note 1 or 2, further including an update count totalizing unit that counts the number of times the first speech recognition dictionary has been updated for each of the set of words.
[0053]
(Supplementary note 4) The speech recognition dictionary adjustment device according to supplementary note 1 or 2, wherein the first speech recognition dictionary is stored for each user.
[0054]
(Supplementary note 5) The speech recognition dictionary adjustment device according to Supplementary note 1 or 2, wherein the first speech recognition dictionary is stored for each environment used by the user.
[0055]
(Supplementary note 6) The speech recognition dictionary adjustment device according to Supplementary note 1 or 2, wherein a plurality of the second speech recognition engines are used.
[0056]
(Supplementary Note 7) A first step of inputting a voice signal emitted by the user;
A second step of recognizing the input voice signal using a first voice recognition dictionary configured in a group of words;
A third step of extracting the generation time of the voice signal and the voice signal in the unit of words that have been at least the recognition target from the recognition result in the second step,
For each occurrence time, a fourth step of re-recognizing using a second dictionary for speech recognition configured on a word-by-word basis using the corresponding speech signal;
A fifth step of storing, based on the recognition result in the fourth step, the group of word-unit speech signals and corresponding reading information as a pair of data, a fifth step of speech dictionary adjustment. Method.
[0057]
(Supplementary Note 8) A first processing step of inputting a voice signal emitted by a user;
A second processing step of recognizing the input voice signal by using a first voice recognition dictionary configured in a unit of words;
A third processing step of extracting the generation time of the voice signal and the voice signal in the unit of words that have been at least the recognition target from the recognition result in the second processing step,
For each occurrence time, a fourth processing step of re-recognizing using the second speech recognition dictionary configured for each word using the corresponding speech signal,
And a fifth processing step of storing, based on the recognition result in the fourth processing step, the set of speech signals in units of words and corresponding reading information as a pair of data. A computer-executable program characterized by the following:
[0058]
【The invention's effect】
As described above, according to the dictionary adjustment device for speech recognition according to the present invention, speech data is extracted in units of words by performing speech recognition processing in units of words, and then speech data in units of words is extracted for speech data corresponding to the words. By re-recognizing, unregistered syllables and phonemes do not occur, and words used by the user can be reliably added to the dictionary for speech recognition. In addition, it is possible to make an adjustment so as to be an efficient speech recognition dictionary with high speech recognition accuracy.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of a speech recognition dictionary adjustment device according to an embodiment of the present invention; FIG. 2 is a data configuration of a first speech recognition dictionary in the speech recognition dictionary adjustment device according to the embodiment of the present invention; FIG. 3 is a diagram illustrating a data configuration example of a second speech recognition dictionary in the speech recognition dictionary adjustment device according to the embodiment of the present invention. FIG. 4 is a speech recognition dictionary adjustment according to the embodiment of the present invention. FIG. 5 is a diagram illustrating an example of a data configuration of a recognition result storage unit in the apparatus. FIG. 5 is another configuration diagram of a speech recognition dictionary adjustment device according to an embodiment of the present invention. FIG. 6 is a speech recognition dictionary according to an embodiment of the present invention. FIG. 7 is a diagram illustrating an example of a data configuration of a first dictionary for speech recognition and a recognition result storage unit in the adjustment device. FIG. 7 is a diagram illustrating a first dictionary for speech recognition and storage of recognition results in the speech recognition dictionary adjustment device according to the embodiment of the present invention. Data structure of division FIG. 8 is another configuration diagram of the speech recognition dictionary adjustment device according to the embodiment of the present invention. FIG. 9 is a flowchart of processing in the speech recognition dictionary adjustment device according to the embodiment of the present invention. ] Illustration of computer environment [Explanation of reference numerals]
11 User 12 First Speech Recognition Engine 13 First Speech Recognition Dictionary 14 Recognition Result Extraction Unit 15 Second Speech Recognition Engine 16 Second Speech Recognition Dictionary 17 Recognition Result Storage Unit 18 Recognition Dictionary Update Unit 19 3 voice recognition engine 20 user ID recognition unit 101 line storage device 102 portable recording medium 102-1 such as CD-ROM or flexible disk CD-ROM
102-2 Flexible disk 103 Computer 104 Recording medium such as RAM / hard disk on computer

Claims (6)

An audio signal input unit for inputting an audio signal emitted by the user;
A first voice recognition unit that recognizes the input voice signal using a first voice recognition dictionary configured in a unit of words;
An audio signal information extraction unit that extracts an audio signal in units of words and a generation time of the audio signal, which are at least recognition targets, from a recognition result in the first audio recognition unit;
For each occurrence time, a second speech recognition unit that recognizes again using a second speech recognition dictionary configured for each word using the corresponding speech signal,
A recognition result storage unit for storing the speech signal in units of words and corresponding reading information as a pair of data based on a recognition result in the second speech recognition unit. Dictionary adjustment device. The apparatus for adjusting a dictionary for speech recognition according to claim 1, further comprising a recognition dictionary updating unit that updates the first dictionary for speech recognition using the stored pair of data. The speech recognition dictionary adjustment device according to claim 1, wherein the first speech recognition dictionary is stored for each user. 3. The speech recognition dictionary adjusting device according to claim 1, wherein a plurality of said second speech recognition engines are used. A first step of inputting a voice signal emitted by the user,
A second step of recognizing the input voice signal using a first voice recognition dictionary configured in a group of words;
A third step of extracting the generation time of the voice signal and the voice signal in the unit of words that have been at least the recognition target from the recognition result in the second step,
For each occurrence time, a fourth step of re-recognizing using a second dictionary for speech recognition configured on a word-by-word basis using the corresponding speech signal;
A fifth step of storing, based on the recognition result in the fourth step, the group of word-unit speech signals and corresponding reading information as a pair of data, a fifth step of speech dictionary adjustment. Method. A first processing step of inputting an audio signal emitted by the user;
A second processing step of recognizing the input voice signal by using a first voice recognition dictionary configured in a unit of words;
A third processing step of extracting the generation time of the voice signal and the voice signal in the unit of words that have been at least the recognition target from the recognition result in the second processing step,
For each occurrence time, a fourth processing step of re-recognizing using the second speech recognition dictionary configured for each word using the corresponding speech signal,
And a fifth processing step of storing, based on the recognition result in the fourth processing step, the set of speech signals in units of words and corresponding reading information as a pair of data. A computer-executable program characterized by the following:

Images