JPH08110790A - Sound recognizing device - Google Patents

Abstract

PURPOSE: To provide a word sound recognizing device capable of coping with the unspecified speaker sound recognition without requiring a registering means other than a sound input apparatus even when the sound registration is limited to once. CONSTITUTION: The word sound is converted into a character train by the unspecified speaker sound recognition of a sound recognition section 2 and registered in a word dictionary 5 in advance. The word sound input is converted into a character train in the similar procedure at the time of recognition, this character train is compared with the character trains registered in the word dictionary 5, and the character train having the highest similarity is outputted from an indicator 8 and an output terminal 9 as the recognition result.

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 for creating a word dictionary of character strings or voice patterns by learning and recognizing an input voice using this word dictionary.

[0002]

2. Description of the Related Art Although research and development relating to voice recognition have been actively carried out in the past, there are few successful commercial systems. One of the causes is that high recognition performance cannot be obtained despite the troublesomeness of registering the recognition voice dictionary. There are two methods for generating a voice dictionary, that is, a template. One is a method that uses a pattern acquired by the same method as the input voice pattern as it is as a voice dictionary (structural pattern matching method), and the other is a method that is acquired by the same method as the input voice pattern. This is a method (statistical pattern matching method) of collecting a large number of the generated patterns to generate a statistical voice model and comparing whether or not the input voice pattern matches this voice model. A general example of the structural pattern matching method is a DP matching method, and a statistical pattern matching method is a recognition method using an HMM (Hidden Markov Model) or a neural network.

The advantage of the structural pattern matching method is that the input voice pattern can be used as it is in the recognition dictionary. Therefore, the recognition device can be used for the voice (the word voice) immediately after the voice (the word voice) to be recognized is registered once. However, since the quality of the registered voice pattern is directly reflected in the recognition performance, it is necessary to pay close attention to the voice during registration, and if the voice registration fails, the recognition rate is drastically reduced. In addition, the speech dictionary that considers the extension to the unspecified speaker speech recognition becomes large.

On the contrary, in the statistical pattern matching method, most voice variations (temporal, frequency characteristic, time-dependent change) are obtained with a small number of voice patterns by inputting the voice to be recognized many times and statistically learning. Since it can deal with fluctuations, it maintains high recognition performance and can easily be extended to unspecified speaker speech recognition. However, it is not possible to obtain high recognition performance simply by registering the voice to be recognized a few times.

Therefore, as a technique which is easy for the user to familiarize with, a voice registration procedure for obtaining a recognition performance similar to that of the statistical pattern matching method by a simple voice registration operation such as the structural pattern matching method is desired. That is, based on the statistical pattern matching method, a voice recognition device that can be used immediately by one registration operation is required. An example of such a speech is as described in the literature, Hirayama, Yoshida, Hattori, “Unspecified speaker word recognition by DSP”, Proceedings of the 1994 IEICE Spring Conference, lecture number A-261. There is a recognition device. This device already has a built-in unspecified speaker voice recognition dictionary in semi-syllable units based on the statistical pattern matching method. When a label name corresponding to a voice pattern to be recognized, for example, a word is input as a character string, , An unspecified speaker voice pattern is automatically generated from a character string, and this is used as a word voice pattern. In such a voice recognition device, an unspecified speaker voice recognition dictionary with high recognition accuracy can be constructed by only inputting a character string once.

[0006]

However, when a device that does not have a function of inputting a character string, for example, a very small voice recognition device is configured, or a recognition vocabulary is registered from a remote place via a telephone line. It is difficult to apply this technology to services such as. By the way, if the character string itself, which is the source of the recognition speech pattern, can be automatically generated by the result of unspecified speaker speech recognition in syllable, phonological unit or a time unit shorter than that, from the character string described above, It becomes possible to apply a voice recognition technology that automatically generates a voice pattern. That is, the voice recognition function can be used by one-time voice registration, and high recognition performance can be obtained for an unspecified speaker.

[0007] At this time, each of the speech units when generating a speech pattern from a character string can be shared as a speech dictionary used in speech recognition in syllable and phonological units. Further, in this case, the recognition dictionary can be configured by the character string itself instead of the voice pattern. Furthermore, when the character string resulting from syllable and phonological recognition at the time of voice registration is a notation that can be visually or auditorily understood by the speaker, that is, for example, "nakadai", or "na", "ka" ,
When the result of voice recognition of "da" and "i" is output as a character string that can be written, for example, "nakadai" or "NAKADAI", the result of syllable recognition is used. When a word voice pattern “Nakadai” is created and used for word voice recognition, the speaker may visually or auditorily confirm the recognition result as a character string “Nakadai” or “NAKADAI”. It will be possible.

An object of the present invention is to automatically generate a voice dictionary for word voice recognition from a voice recognition result of an unspecified speaker having a syllable or a phonological level, so that high recognition performance can be obtained even if voice registration is limited to one time. An object of the present invention is to provide a word voice recognition device which can be obtained and can deal with an unspecified speaker, and which does not require voice registration means other than a voice input device.

[0009]

According to the invention of claim 1, a means for inputting a voice, a means for recognizing a syllable or a phoneme level for the input voice, and a voice dictionary used for the recognition. And a means for converting the recognition result into a character string, a means for saving the converted character string as a word dictionary, and a means for comparing the character string saved in the word dictionary with the input character string. It is composed of means for outputting the result.

According to the invention of claim 2, in addition to the invention of claim 1, means for further generating a voice pattern for word recognition from the word dictionary and the voice dictionary, and storing the generated voice pattern as a dictionary. Means for recognizing a word level between a voice pattern stored in a dictionary and an input voice, and a means for outputting a result of word voice recognition instead of a means for outputting a result. It is composed of

According to the invention of claim 3, to the invention of claim 1 or 2, means for presenting a character string output as a part of the recognition result to the speaker is added.

[0012]

According to the first aspect of the invention, the isolated uttered word voice is converted into a character string by recognizing a syllable or phoneme level unspecified speaker voice, and is converted from a voice into a character string by the same procedure. The content of the word dictionary that is compared with the word dictionary and determined to be the same character string or a similar character string is output as a recognition result.

According to the second aspect of the present invention, the word dictionary for recognition is converted from a voice into a character string in the same manner as in the first aspect of the invention, and further converted from a character string into a word voice pattern. The word speech that is isolated and uttered is collated with this word speech pattern, and the character string that is the generation source of the word speech pattern determined to have the highest similarity is output as the recognition result. It

In either of the first and second aspects of the invention, the speaker can visually or audibly confirm the character string of the voice recognition result.

[0015]

FIG. 1 shows an embodiment based on the invention of claim 1. In FIG. 1, a voice input unit 1 is a device for voice input by a speaker, and is, for example, an audio microphone.
The voice recognition unit 2 performs A / D conversion on the voice input from the voice input unit 1, detects a voice section, analyzes the detected section, and analyzes the unspecified speaker voice in syllable or phoneme unit. It recognizes and outputs a character string as the recognition result. Here, the input voice is a normally spoken word voice such as “nakadai”, or a single voice such as “na”, “ka”, “da”, and “i”. It is a voice that is input by dividing it into syllables. The A / D conversion method is generally performed by a voice recognition method, for example, a sampling frequency is 10 kHz and 12-bit quantization, and the voice section detection method is, for example, a cutout method based on an observed value of short-time power. , The voice analysis method is, for example, L
The PC cepstrum analysis and phonological recognition method is a syllable spotting algorithm for unspecified speaker speech, which is described in, for example, the literature, Nakagawa, “Speech recognition by probabilistic model”, IEICE. In the voice dictionary unit 3, a standard speech pattern of a syllable or a phoneme used in the voice recognition unit 2 is given a label name corresponding to it (for example, a label name “A” for a pattern of the voice “A”). This voice pattern is prepared by, for example, HMM which is an example of the statistical pattern matching method shown in the above-mentioned document. The language processing unit 4 prepares the character string of the recognition result output from the voice recognition unit 2 into a word, and sends it to the word dictionary unit 5 and the comparison unit 6. The output of the voice recognition unit 2 is a word voice (continuous syllable) in which the input voice is “Nakadai”
In this case, addition of unnecessary syllables and omission of syllables are observed only with the syllable recognition result of the first place. Therefore, a form in which a plurality of recognition results are recognized for one phoneme section, that is, for example, phonological lattice It may be output in the format. Therefore, the language processing unit 4 plays a role of organizing the output character string of the voice recognition unit 2 by adding language information. The word dictionary unit 5 stores the output word (character string) of the language processing unit 4 as a dictionary for word recognition. The comparison unit 6 compares the character strings between the words accumulated in the word dictionary unit 5 and the output of the language processing unit 4, and outputs the result to the display 8 and the output terminal 9. The switch 7 switches the operation of the comparison unit 6 between learning and recognition. The display 8 displays the recognition result to the speaker, and is, for example, a character display or a voice synthesizer. The output terminal 9 outputs the final recognition result to an external device, and is, for example, an interface unit of a computer.

The operation of the word voice recognition apparatus shown in FIG. 1 will be described below. This operation is divided into two operations: registration and recognition. (1-1) Case of Registration At the time of registration, this device performs an operation of registering the input word voice as a word (character string) in the word dictionary unit 5.
The speaker first switches the comparison unit 6 to the operation at the time of registration with the switch 7, and then outputs a voice that can be recognized in syllable units to the voice input unit 1. The voice recognition unit 2 recognizes the unspecified speaker voice in syllable units, converts it into a word (character string) in the language processing unit 4, and sends it to the comparison unit 6. The comparison unit 6 compares the input word with a plurality of words (character strings) that have been converted from speech into character strings and stored in the word dictionary unit 5 in the same manner as above. Here, if there is a word that is judged to be the same or very similar between the input word and the word stored in the word dictionary unit 5, the comparison unit 6 causes the current word voice input. It is determined that the contents cannot be registered in the word dictionary unit 5, the input word is discarded, information indicating that the word cannot be registered is displayed on the display unit 8 and the speaker is prompted to input the voice again. That is, the utterer usually knows a word already registered in the word dictionary unit 5, and when it is determined that the word is the same as the vocabulary, it is determined that the utterer's utterance was bad, and the utterer inputs again. On the other hand, when the input word is different from all the words stored in the word dictionary unit 5, the comparison unit 6 registers the input word (character string) in the word dictionary unit 5 as a new word, and the registration is completed. The display 8 notifies the speaker of the completion. here,
When the input and recognized word (character string) is a notation that can be visually or auditorily understood by the speaker, that is, the result of recognizing by inputting “Nakadai” is Dai ”or“ Nakadai ”
In such a case, the display device 8 can display a character string as a registration result. (1-2) At the time of recognition At the time of recognition, this device recognizes the input word voice and outputs the result to the output terminal 9. The speaker first switches the comparison unit 6 to the recognition operation by the switch 7,
Next, a word voice is uttered to the voice input unit 1. As in the case of registration, this voice is recognized by the voice recognition unit 2 as a syllable-based unspecified speaker voice, converted into words (character strings) by the language processing unit 4, and sent to the comparison unit 6. The comparing unit 6 compares the input word with a plurality of words (character strings) converted from speech to character strings and stored in the word dictionary unit 5 by the same method as above. Here, if there is a word that is determined to be the same or very similar between the input word and the word stored in the word dictionary unit 5, the comparison unit 6 determines that word as the recognition result. Output from the output terminal 9. Further, as in the case of registration, when the word (character string) which is the recognition result is a notation that can be visually or aurally understood by the speaker, the display 8 simultaneously displays the output end 9 and the registration result. It is possible to display a character string as.

Next, a second embodiment of the present invention will be described with reference to FIG. In FIG. 2, the voice input unit 1, the voice recognition unit 2, the voice dictionary unit 3, the language processing unit 4, the word dictionary unit 5, the comparison unit 6, the display unit 8, and the output terminal 9 are assigned the same numbers in FIG. It has the same function as the above-mentioned part. However,
In FIG. 2, the voice pattern registered in the voice dictionary unit 3 is also used in the word voice generation unit 10 to generate a word voice pattern, and the output end 9 is not the output from the comparison unit 6 but the word The recognition result of the voice recognition unit 12 is output. The switch 7 switches the output destination of the voice input unit 1 between learning and recognition. The word voice generation unit 10 generates a word voice pattern of an unspecified speaker for word voice recognition from the words (character strings) accumulated in the word dictionary unit 5 and the voice pattern accumulated in the voice dictionary unit 3. The word voice dictionary unit 11 stores the word voice patterns of the unspecified speaker generated by the word voice generation unit 10. The word voice recognition unit 12 performs A / D conversion on the input word voice, detects a voice section, analyzes the detected voice section, and analyzes a plurality of words stored in the word voice dictionary section 11. Speech recognition is performed on a word-by-word basis with the speech pattern. A /
The D conversion, voice section detection, voice analysis method and the like are the same as those of the voice recognition unit.

The operation of this embodiment shown in FIG. 2 will be described below. This operation is also divided into two, that is, at the time of registration and at the time of recognition, as in the embodiment shown in FIG. (2-1) Case of Registration At the time of registration, this device performs an operation of converting an input word voice into a word (character string) and further converting into a word voice pattern. The speaker first switches the output destination of the voice input unit 1 to the voice recognition unit 2 with the switch 7, and then speaks a word voice to the voice input unit 1. Similar to the embodiment shown in FIG. 1, this voice is recognized by the voice recognition unit 2 in the syllable unit as an unspecified speaker voice, converted into a word (character string) by the language processing unit 4, and sent to the comparison unit 6. The comparing unit 6 compares the input word with a plurality of words (character strings) converted from speech to character strings and stored in the word dictionary unit 5 by the same method as above. Here, if there is a word that is determined to be the same or very similar between the input word and the word stored in the word dictionary unit 5, the comparison unit 6 determines that the current word voice input content is present. Judges that the word cannot be registered in the word dictionary unit 5, discards the input word, displays information indicating that the word cannot be registered on the display 8 and prompts the speaker to input the voice again. Further, when the input word is different from all the words accumulated in the word dictionary unit 5, the comparison unit 6 registers the input word (character string) in the word dictionary unit 5 as a new word,
The display unit 8 notifies the speaker that registration has been completed. Here, as in the case of the first embodiment, when the input and recognized word (character string) is a notation that can be visually or auditorily understood by the speaker, the display device 8 is registered. As a result, it is possible to display a character string. After that, the word dictionary unit 5 sends the newly registered word (character string) data to the word voice generation unit 10. The word voice generation unit 10 generates a word voice pattern of an unspecified speaker from a character string based on the voice pattern registered in the voice dictionary unit 3. The generated word voice pattern is a concatenation of syllable voice patterns, but it is also possible to perform pattern shaping (concatenation learning) as a word voice after the concatenation. The word voice pattern generated in this manner is given a word (character string) label that is the source of the pattern generation, and is accumulated in the word voice dictionary unit 11. (2-2) At the time of recognition At the time of recognition, this device recognizes the input word voice and outputs the result to the output terminal 9. The speaker first switches the output destination of the voice input unit 1 by the switch 7 to the word voice recognition unit 1
2 is switched to, and then a word voice is uttered to the voice input unit 1. This voice is subjected to voice analysis in the word voice recognition unit 12 similarly to the operation of the voice recognition unit 2, and is subjected to pattern matching with a plurality of word voice patterns accumulated in the word voice dictionary unit 11 on a word-by-word basis. As a result, the label assigned to the word voice pattern determined to have the highest degree of similarity is output from the output terminal 9 as the word voice recognition result.
Here, as in the case of the first embodiment, when the label (character string) given to the word voice pattern of the recognition result is a notation that can be visually or auditorily understood by the speaker, The display 8 can display a character string as a registration result at the same time as the output terminal 9.

1 and 2, parts other than the voice dictionary unit 3, the word dictionary unit 5, and the word voice dictionary unit 11 that perform various processes and analysis are performed by an electronic computer or a DSP (digital signal processor). Each unit does not have to constitute independent hardware.

[0020]

As described above, according to the present invention, in the word voice recognition device using the statistical pattern matching method, the word voice pattern for the recognition dictionary is obtained from the unspecified voice recognition result of the syllable or the phoneme level. By automatically generating as a character string or a voice pattern generated from the character string, (1) it can be used as a word voice recognition device for a specific person and for an unspecified speaker with a single voice input. , (2) It is possible to create a word voice pattern that was conventionally performed by inputting characters with a character input device. (3) Unspecified voice recognition of syllable / phoneme level during recognition dictionary generation When the character string generated by is a notation that can be visually or auditorily understood by the speaker, a word voice pattern is generated as a recognition result. The string is then presented to the speaker, it is possible to check, having an effect.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a block diagram showing a second embodiment of the present invention.

Claims (3)

[Claims] 1. A means for inputting a voice, a means for recognizing a syllable or a phoneme level for the inputted voice, a voice dictionary for storing a voice pattern of a syllable or a phoneme level, and a dictionary used for the recognition. , A means for converting the recognition result into a character string, a means for storing the converted character string as a word dictionary, the character string stored in the word dictionary and the voice input, and the character of the recognized result It is composed of a means for comparing with a string and a means for outputting the result of the comparison, and converts a syllable or a phonological level to an isolated uttered word voice to convert it into a character string, and the converted character string A voice recognition device comprising means for comparing the contents of the word dictionary with the contents of the word dictionary and outputting the result. 2. A means for generating a voice pattern for word recognition from a character string stored in the word dictionary and a voice pattern stored in the voice dictionary, and the generated voice pattern and a generation source. A word-speech dictionary that stores a combination of character strings as a dictionary, a means for recognizing a word-level between a speech pattern stored in the word-speech dictionary and an input speech, and a word-speech isolated Recognizing the syllable or phonological level to convert to a character string, to generate the word voice dictionary for word voice recognition from the converted character string, the generated word voice dictionary, and the newly input word The voice recognition device according to claim 1, further comprising means for performing recognition with a voice and outputting the result. 3. The voice recognition device according to claim 1, further comprising means for presenting a character string output as a part of the recognition result to a speaker.