JPH11202890A - Speech retrieval device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech retrieval device which can accurately retrieve speech data from a speech data base. SOLUTION: This device has a speech data base 1 stored with speech data to be retrieved, a 1st speech recognition part 2 which detects specific speech data by spotting among the speech data of the speech data base 1, a 2nd speech recognition part 3 which recognizes a key word from the voice that a user has spoken, and a speech dictionary generation part 4 which generates a speech dictionary for the 1st speech recognition part 2 from the key word recognized by the 2nd speech recognition part 3, and the 1st speech recognition part 2 uses the speech dictionary generated by the speech dictionary generation part 4 and detects speech data corresponding to the speech dictionary by spotting among the speech data of the speech data base.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice search device for searching the contents of a voice database.

[0002]

2. Description of the Related Art In order to search for a keyword for voice recognition from voice data such as an answering machine or voice mail, or voice data accompanying a video such as a video, conventionally,
2. Description of the Related Art A voice search device as disclosed in JP-A-6-175698 is known. In this voice search device, as a first method, when a user utters a keyword, words corresponding to the keyword uttered by the user are sequentially detected from voice data in a voice database by word spotting (recognition). There is a method of retrieving voice data intended by a user from a voice database. Alternatively, as a second method, the voice database has a voice data storage unit and a keyword storage unit, and voice data such as an answering machine or an e-mail is stored in the voice data storage unit, and further a keyword storage is performed. When a user utters a keyword when a keyword for the voice data in the voice data storage unit is stored in advance in the voice data storage unit, the keyword uttered by the user is stored in the keyword storage unit of the voice database (keyword candidate). )
When a matching keyword is detected as a result of the matching, there is a method of extracting voice data (voice signal) in the voice data storage unit corresponding to the keyword.

[0003]

However, in the case of the first method, since the user and the speaker (creator) of the voice database are usually different (these voices are usually
There is a drawback that recognition performance (search performance) is not good. In particular, if the gender of the speaker is different, such as when the user is a woman and the voice database is news by a male announcer, it will be difficult to recognize accurately.

In the second method, since the voice of the speaker is not directly compared with the voice of the voice database, the above-mentioned disadvantage of the first method is not present, but the keyword must be stored in the voice database in advance. Disadvantages arise.
This is not simply a matter of inputting time, but it is necessary to predict keywords (search tags) that the user will need in advance and make them as keyword candidates. Otherwise, there will be no keyword required for searching.

[0005] It is an object of the present invention to provide a voice search device capable of accurately searching voice data in a voice database.

[0006]

In order to achieve the above object, according to the first aspect of the present invention, there is provided a voice database storing voice data to be searched, and a voice database stored in the voice database. A first voice recognition unit for detecting voice data by spotting, and a second voice recognition unit for recognizing a keyword from voice uttered by a user.
And a speech dictionary creating unit that creates a speech dictionary for the first speech recognition unit from the keywords recognized by the second speech recognition unit. The first speech recognition unit ,
The voice dictionary corresponding to the voice dictionary is detected by spotting from the voice data of the voice database using the voice dictionary created by the voice dictionary creating unit.

According to a second aspect of the present invention, in the voice search device according to the first aspect, keyword storage means for storing keyword candidates is further provided, and the second voice recognition unit is used. The keyword candidate is recognized from the voice uttered by the person, and the keyword candidate is stored in the keyword storage unit. The voice dictionary creating unit spots the keyword candidate stored in the keyword storage unit, and determines the first keyword from the spotted keyword. Is characterized in that a voice dictionary for the voice recognition unit is created.

According to a third aspect of the present invention, in the voice search device according to the second aspect, the keyword candidates stored in the keyword storage means have a hierarchical structure of at least two levels of a large category and a small category. Described in grammar, the speech dictionary creation unit creates a speech dictionary for the first speech recognition unit from the recognition result of the keyword described in the item of the small classification.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a voice search device according to the present invention. Referring to FIG. 1, the voice search apparatus includes a voice database 1 storing voice data to be searched, and a voice database 1 for detecting predetermined voice data from voice data of the voice database 1 by spotting. 1 voice recognition unit 2, a second voice recognition unit 3 for recognizing a keyword from voice uttered by the user, and a first voice recognition unit 2 based on the keyword recognized by the second voice recognition unit 3. And a voice dictionary creating unit 4 for creating a voice dictionary for
The audio dictionary corresponding to the audio dictionary is detected by spotting from the audio data in the audio database using the audio dictionary created by the audio dictionary creating unit 4.

FIG. 2 is a diagram showing a specific example of the second speech recognition section 3. In the example of FIG. 2, the second speech recognition unit 3
The voice recognition is performed using the dictionary for the voice recognition unit. This dictionary contains words of interest to users
(Groups), for example, voice standard patterns such as "dog", "cat", "japan",... Are registered in advance. in this case,
When the user utters “Inu” and this voice is input to the second voice recognition unit 3, the second voice recognition unit 3 extracts the feature pattern of this voice “Inu” and stores it in the dictionary. It is collated with a voice standard pattern, and the highest similarity is output as a recognition result.

In the above-described example, the dictionary is created for each word, but may be created for each phoneme. Also, it may be for a specific speaker or for an unspecified speaker.

FIG. 3 is a diagram showing an example of the configuration of the speech dictionary creation unit 4. In this example, the speech dictionary creation unit 4 performs the kana-phoneme conversion table 11 and the recognition by the second speech recognition unit 3. When the resulting character string (kana) is input, a phoneme sequence corresponding to this character string (kana) is obtained from the kana-phoneme conversion table 11-
When a phoneme sequence is given from the phoneme conversion unit 12, the phoneme dictionary 13, and the kana-phoneme conversion unit 12, the phoneme sequence is converted into a parameter sequence by referring to the phoneme dictionary 13. And a voice dictionary conversion unit 14.

FIG. 4 is a diagram showing an example of a kana-phoneme conversion table 11, in which phonemes / a /, / i /,... Corresponding to kana "a", "i",. . In this case, when the kana-phoneme conversion unit 12 inputs, for example, "Inu" as a kana character string, the kana-phoneme conversion table 11 uses the kana-phoneme conversion table 11 of FIG. Is converted into a phoneme sequence.

In FIG. 3, the phoneme dictionary 13 includes
Voice parameters for unspecified speakers are registered in advance.
That is, in the phoneme dictionary 13, feature vectors (usually plural) and parameters of duration are registered for each phoneme. Also, the voice dictionary conversion unit 14 outputs, for example, / i /, / n /, / u / as a phoneme sequence from the kana-phoneme conversion unit 12.
Is output, the parameters in the phoneme dictionary 13 are arranged for each phoneme in the order of the phoneme sequence (/ i /, / n /, / u /) to create a speech dictionary. In the above example,
Since the phoneme sequence (/ i /, / n /, / u /) is the phoneme sequence of the word “Inu”, the speech dictionary created as described above is a word dictionary.

FIG. 5 is a diagram showing a configuration example of a speech dictionary created by the speech dictionary creating section 4. In the example of FIG. 5, one recognition result, for example, “Nihon” is directly input as a character string to the voice dictionary creating unit 4 of FIG. 5 (ie, “Nihon”).
Input as ").

Next, the processing operation of the voice search device of FIG. 1 will be described. When the user utters a keyword and inputs the user's voice (keyword) from a voice input unit (not shown) such as a microphone, the input voice is input.
The (keyword) is converted into characteristic parameters required for speech recognition, and speech recognition is performed by the second speech recognition unit 3. Here, any method can be used for the feature parameter and the matching method of the voice recognition in the second voice recognition unit 3. For example, a method described in the document "" Speech Recognition by Stochastic Model "(by Seiichi Nakagawa, IEICE) can be used.

However, since the second voice recognition section 3 only needs to have a function of recognizing the user's voice, the second voice recognition section 3 recognizes the user's voice most accurately or at high speed. Any device that can be recognized can be used. For example,
It is possible to use a specific speaker type speech recognition device using a phoneme model trained only with the user's speech data.
Alternatively, an unspecified speaker-type recognition device having a speaker adaptation function can be selected and a device that is sufficiently adapted to the user's voice can be used.

The keyword (usually expressed by kana characters) recognized by the second voice recognition unit 3 is converted into a voice dictionary for the first voice recognition unit 2 by the voice dictionary creation unit 4. . As a method for creating the speech dictionary, a speech recognition method that is not a word-based recognition method, for example, a method generally used in a continuous speech recognition method or a large vocabulary speech recognition in units of phonemes can be used. In a method that is not such a word-based method, as described above, a phoneme dictionary is provided, and a speech dictionary (word dictionary), that is, a dictionary in which phonemes are arranged in order, is created and recognized automatically from the spelling. In this case, it is not necessary to explicitly create a word dictionary (it is not necessary to explicitly create a word dictionary), and the order of phonemes may be specified.

In this method, as in the above-described example, for example, the character string "inu" is used to form / i /, / n /, / u /
Create a voice dictionary called Where / i /, / n /,
Each phoneme of / u / contains an acoustic dictionary suitable for recognition processing in the first speech recognition unit 2 (an unspecified dictionary that does not match the user's voice). Therefore, the first speech recognition unit 2 uses the / i /, / n /, /
The speech database 1 can be searched using the phoneme sequence u / (dictionary of words) as a dictionary.

That is, the speech dictionary created as described above is given to the first speech recognition unit 2, and the first speech recognition unit 2 stores the speech dictionary from the speech dictionary creation unit 4 in the speech database 1. The voice data is collated with the stored voice data, and voice data having a collation score higher than a predetermined threshold is detected. Here, the threshold value may be changed depending on the quality of the audio database (the quality of stored audio data (e.g., whether or not it is clear)) and the required detection accuracy. Therefore, it is preferable that the threshold value be variable. At this time, if the threshold value is set small, a position that does not match the keyword is also detected. Conversely, if the threshold value is large, the position may not be detected even if there is a voice that matches the keyword. Need to be set to something.

As described above, in the voice search apparatus shown in FIG. 1, voice data uttered by the user is not directly compared with voice data in the voice database, so that the voice uttered by the user is not directly used for the search. So) accurate recognition is possible. That is, the second voice recognition unit 3 that recognizes the voice of the user may be a specific speaker type voice recognition device that registers the user's voice or an unspecified speaker type voice recognition device. It is possible to use a speech recognition device that is sufficiently speaker-adapted by the user's voice.
After recognizing the user's voice by the voice recognition unit 3, the voice dictionary for the first voice recognition unit 2 is created from the recognition result. Therefore, when the voice database 1 is recognized (spotted), the voice database 1 is recognized. The best speech recognition device for
Can be used for the voice recognition unit 2. That is, since it is possible to select an optimum recognition device for both recognizing the user's voice and recognizing the voice database 1, it is possible to accurately search for voice data from the voice database 1.

FIG. 6 is a diagram showing a modification of the voice search device according to the present invention. Referring to FIG. 6, the voice search device is the same as the voice search device of FIG. 1, but further includes a keyword storage unit 5 for storing keyword candidates. The unit 3 recognizes the keyword candidates from the voice uttered by the user, stores the keyword candidates in the keyword storage unit 5, and the voice dictionary creating unit 4 spots the keyword candidates stored in the keyword storage unit 5, A voice dictionary for the first voice recognition unit 2 is created from the spotted keywords.

FIG. 7 is a diagram showing an example of the keyword storage means 5. In the example of FIG. 7, the keyword candidates are described in a grammar having a hierarchical structure of at least two or more hierarchies of a major classification and a minor classification. , A speech dictionary for the first speech recognition unit 2 is created.

That is, for example, when searching for animal names, if the number is large, if the second voice recognition unit 3 recognizes them all at once, there are many types of animals (the number of words to be recognized is large).
As a result, the recognition performance may be degraded, and the search may not be performed efficiently. In such a case, as shown in FIG. 7, the keyword candidates are described in a grammar having a structure of a large classification and a small classification, so that the second speech recognition unit 3 can perform stepwise recognition and can perform high-speed processing with high accuracy. Can be recognized. Thus, "country name-city name",
By using a grammar having a hierarchical structure such as "professional name-personal name", the user's voice can be recognized with higher recognition accuracy compared to the case of collectively recognizing keywords in small categories required for search. Spotting becomes possible. Specifically, for example, when searching for a dog named "Pochi", the grammar of the large category "Dog" and the small category "Pochi" is created, and the user speaks the voice "Dog Pochi." When uttering, it is possible to hierarchically recognize, first from this voice, "dog" of a large classification and then "pochi" of a small classification.

Next, the processing operation of the voice search device shown in FIG. 6 will be described. In this voice search device, assuming that the voice uttered by the user (input voice) is, for example, “bird peacock”, the second voice recognition unit 3 targets the entire input voice section in FIG. In the classification table, try to spot large classification items. Then, when the recognition result "bird" is obtained,
Next, for the section other than the section where “bird” was spotted in the input voice section, the spot names “sparrow”, “swallow”, “peacock”,... , "Peacock" can be recognized. Thus 2
By recognizing “peacock” in stages (by performing spotting), the number of target words per recognition can be reduced, and high-speed recognition can be performed with high accuracy.

The speech dictionary creating unit 4 creates a speech dictionary for the "peacock" obtained as a result of the recognition by the second speech recognition unit 3 in this manner, and uses the speech dictionary to create a first speech recognition unit. By performing spotting from the audio database 1 by using 2, it is possible to search for a portion including the audio of "peacock".

As described above, the voice search device of FIG. 6 has the advantage of the voice search device of FIG. 1 (that is, since the voice data uttered by the user is not directly compared with the voice data in the voice database, accurate recognition is not possible). In addition to the above, the keyword candidates are stored in the keyword storage means 5 and the keywords can be recognized by spotting, so that the user does not need to worry about adding unnecessary words, etc. can do.

Further, in the voice search apparatus shown in FIG. 6, as shown in FIG. 7, the keyword candidates stored in the keyword storage means 5 are converted into a grammar having a hierarchical structure of at least two hierarchies of a large classification and a small classification. The voice dictionary creation unit 4 describes
A speech dictionary for the first speech recognition unit 2 is created from the recognition results of the keywords described in the sub-category items. The number of target words per recognition can be reduced, and high-speed recognition can be performed with high accuracy.

In the example described above (the example in FIG. 7), the keyword is classified in two steps. However, the keyword may be classified into more layers.

In each of the above examples, the speech dictionary is
When the second speech recognition unit 3 recognizes, for example, "Nihon", the one speech recognition result is input as a character string to create the speech dictionary. When the second speech recognition unit 3 recognizes, for example, "Japan", the one recognition result "Japan" is input as a first character string, and a similar "Japan" is input to the second character string. May be created by inputting as a character string (inputting two character strings).

That is, in the example of FIG. 8, when the user utters, for example, "Nihon" and this input voice is recognized as "Nihon" by the second voice recognition unit 3, the keyword "Nihon" is obtained. When the voice dictionary is created, the voice dictionary creating unit 4 generates a voice dictionary corresponding to the keyword “Nihon”.
“Nihon” and “Nippon” are created, and in this case, the first speech recognition unit 2 outputs “Nihon” and “Nippon”.
The voice database 1 can be searched (word spotting) using the voice dictionary. Thereby, a more accurate search can be performed.

FIG. 9 is a diagram showing an example of a hardware configuration of the voice search device of FIG. 1 or FIG. Referring to FIG.
This voice search device is realized by, for example, a personal computer or the like, and controls a CPU 21 that controls the entirety.
ROM 22 storing the control program of
RAM 23 used as a work area of PU 21
A voice input device 24 for inputting a user's voice, a data storage device 25 for storing voice data and the like, and a result output device (for example, a display or a printer) 26 for outputting information on a result of voice search. And

Here, the CPU 21 has the functions of the first and second speech recognition units 2 and 3 and the speech dictionary creation unit 4 shown in FIGS.

The functions of the CPU 21 such as the first and second speech recognition units 2 and 3 and the speech dictionary creation unit 4 are, for example, software packages (specifically, information recording such as a CD-ROM). (Medium), and therefore, in the example of FIG.
Is set, the medium drive device 31 that drives this
Is provided.

In other words, the voice search device of the present invention allows a general-purpose computer system having a voice input device, a display, and the like to read a program recorded on an information recording medium such as a CD-ROM, and The present invention can also be implemented in an apparatus configuration in which a microprocessor of the system executes a voice search process. In this case, a program for executing the voice search processing of the present invention (that is, a program used in a hardware system)
Is provided in a state recorded on a medium. The information recording medium on which the program or the like is recorded is not limited to a CD-ROM, but may be a ROM, a RAM, a flexible disk, a memory card, or the like. The program recorded on the medium is installed in a storage device incorporated in the hardware system, for example, a hard disk device, so that the program can be executed to realize the voice search function of the present invention.

[0036]

As described above, according to the first to third aspects of the present invention, a voice database storing voice data to be searched and a voice database stored in the voice database. A first voice recognition unit for detecting predetermined voice data by spotting; and a second voice recognition unit for recognizing a keyword from voice uttered by a user.
And a speech dictionary creating unit that creates a speech dictionary for the first speech recognition unit from the keywords recognized by the second speech recognition unit. The first speech recognition unit ,
Using the voice dictionary created by the voice dictionary creation unit, voice data corresponding to the voice dictionary is detected from the voice data of the voice database by spotting. Since it is possible to select the most appropriate recognition device for both the recognition of the voice data and the voice data, the voice data can be accurately retrieved from the voice database.

In particular, according to the second aspect of the present invention, a keyword storage means for storing keyword candidates is further provided, and the second voice recognition unit converts the keyword candidates from voice uttered by the user. Recognizing and storing the keyword candidate in the keyword storage means,
Since the keyword candidates stored in the keyword storage means are spotted and a voice dictionary for the first voice recognition unit is created from the spotted keywords, the user is concerned about adding unnecessary words. This eliminates the need to do so, and makes it easier to speak.

According to the third aspect of the invention, the keyword candidates stored in the keyword storage means are described in a grammar having a hierarchical structure of at least two hierarchies of a large classification and a small classification, and the speech dictionary creation unit Creates a speech dictionary for the first speech recognition unit from the recognition results of the keywords described in the items of the small classification, so that the number of target words per recognition can be reduced, and Good and fast recognition.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration example of a voice search device according to the present invention.

FIG. 2 is a diagram illustrating a specific example of a second speech recognition unit.

FIG. 3 is a diagram illustrating a configuration example of a voice dictionary creation unit.

FIG. 4 is a diagram showing an example of a kana-phoneme conversion table.

FIG. 5 is a diagram illustrating an example of a speech dictionary created by a speech dictionary creation unit.

FIG. 6 is a diagram showing a configuration example of a voice search device according to the present invention.

FIG. 7 is a diagram illustrating an example of a keyword storage unit.

FIG. 8 is a diagram illustrating another example of the speech dictionary created by the speech dictionary creation unit.

FIG. 9 is a diagram illustrating an example of a hardware configuration of the voice search device in FIG. 1 or FIG. 6;

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Speech database 2 1st speech recognition part 3 2nd speech recognition part 4 Speech dictionary creation part 5 Keyword storage means 21 CPU 22 ROM 23 RAM 24 Speech input device 25 Data storage device 26 Result output device 30 Information storage medium 31 Medium Drive

Claims (3)

[Claims] An audio database in which audio data to be searched is stored; a first audio recognition unit for detecting predetermined audio data from audio data in the audio database by spotting; A second speech recognition unit for recognizing a keyword from the voice uttered by the user, and a first speech recognition unit based on the keyword recognized by the second speech recognition unit.
And a voice dictionary creating unit for creating a voice dictionary for the voice recognition unit. The first voice recognition unit uses the voice dictionary created by the voice dictionary creating unit to convert the voice data of the voice database. A voice search device characterized by detecting voice data corresponding to a voice dictionary from inside by spotting. 2. The voice search device according to claim 1, wherein
Keyword storage means for storing keyword candidates is further provided, and the second voice recognition unit recognizes keyword candidates from voice uttered by the user, stores the keyword candidates in the keyword storage means, The voice dictionary creating unit spots keyword candidates stored in the keyword storage unit and creates a voice dictionary for a first voice recognition unit from the spotted keywords. Search device. 3. The voice search device according to claim 2, wherein
The keyword candidates stored in the keyword storage means are described in a grammar having a hierarchical structure of at least two levels of a large classification and a small classification, and the speech dictionary creating unit recognizes the keyword described in the item of the small classification. A speech search device for creating a speech dictionary for a first speech recognition unit from a result.