JP2002372988A - Recognition dictionary preparing device and rejection dictionary and rejection dictionary generating method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition dictionary for rejecting a noise or the utterance of words other than recognition words according to the use form of voice recognition using discrete word recognition, and for returning correct voice recognized results only when the recognition words are uttered, and to provide a framework for discriminatingly using rejection vocabulary due to a noise and rejection vocabulary due to the utterance of a user from a constructed rejection dictionary. SOLUTION: The group of rejection words to be rejected is stored in a rejection word data base 5. As for the rejection words to be actually used for voice recognition, recognition words stored in a recognition word storing part 41 and recognized results obtained at the time of erroneously recognizing the correct voices of recognition words inputted by a correct voice inputting part 8 as rejection words are removed from the rejection words in the rejection word data base 5 including similar words generated by a similar word generating part 7 by a rejection word removing part 6, and the words included in a noise rejection word data base are stored in a noise rejection word storing part 422, and the other words are stored in an utterance rejection word storing part 421. At the time of using voice recognition, voice recognition is operated to the dictionary including the words stored in the recognition word storing part 41, the utterance rejection word storing part 421, and the noise rejection word storing part 422.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for rejecting ambient noise such as mechanical noise or speech and speech outside the grammar in a speech recognition system.

[0002]

2. Description of the Related Art A major problem in a system using speech recognition is that noise or surrounding speech is erroneously recognized as a recognized word, resulting in malfunction. Also,
In addition to the noise, there is a problem that the user may similarly misrecognize when the user speaks a vocabulary other than the recognized word.

As a method of preventing such erroneous recognition, there is a method of rejecting noise as speech other than the recognized word. As a conventional noise rejection method, Japanese Patent Application Laid-Open No. H11-288295 discloses a word dictionary unit in which a plurality of recognition target words of one utterance are registered, and a word storage unit in which word data to be recognized is registered, It is composed of a noise word storage unit in which noise word data close to noise is registered.

[0004] The word data is a word group in which a plurality of words to be recognized are registered, and the same word group as words actually uttered is registered. The noise word data is a word group in which a plurality of noise words close to the noise to be removed are registered, and a plurality of words that are likely to be erroneously recognized by the noise are registered. In other words, words that are selected by using noise once are used, and necessary types of words are registered according to their voice environment and similarity with the word data 9 to be recognized.

[0005] For example, a single syllable such as "A" or "U" and words "Ah" or "Au" obtained by combining the two syllables are registered in the word storage unit as word data, and true noise, for example, Recognition of noises such as noises of turning over paper, sounds of people walking, and the like are performed, and the recognized words are picked up and registered in the noise word storage unit 3 as noise word data.

On the other hand, Japanese Unexamined Patent Publication No. 2-89100 and the like describe noise mixed into a microphone and reference to a recognition word registered in a dictionary.

[0007]

However, in order to recognize a word spoken by an unspecified speaker to an robot from an undefined position like a robot, a method of simply registering noise in advance as a recognition word and rejecting the noise is as follows. There were cases where the vocabulary that could be registered was limited and sufficient rejection performance could not be obtained. In addition, it was not possible to cope with erroneous recognition when the user uttered speech other than the recognized word.

According to the present invention, an enormous rejection dictionary using a syllable net is constructed in advance, and only rejection words that affect recognition words used for speech recognition are removed, thereby making it possible to generate utterances other than noise and recognition words. An object of the present invention is to construct a speech recognition dictionary that rejects and only returns a correct speech recognition result when a recognized word is uttered. Another object of the present invention is to provide a framework for distinguishing and using a rejected vocabulary due to noise and a rejected vocabulary based on a user's utterance from the constructed rejection dictionary.

[0009] Further, it is expected that a speech recognition dictionary including a noise rejection vocabulary adapted to a use form of speech recognition can be constructed by using the present method.

[0010]

In a system for inputting human voice and returning a voice recognition result, a voice input unit for receiving a human voice utterance and outputting a voice input signal, and voice recognition based on the voice input signal. A speech recognition unit that outputs results in text format, a speech output unit that outputs speech recognition results to an external output device or another program, a recognition word storage unit that stores recognition words, and a rejection of invalid speech utterance or noise Recognition dictionary unit used for speech recognition in the speech recognition unit, a rejection word database that stores rejection words comprehensively, and recognition passed from the recognition word storage unit. A rejection word removing unit that removes the vocabulary matching the vocabulary from the rejection word database and stores the remaining vocabulary in the rejection word storage unit. A noise rejection word database consisting of a rejection word storage unit and a noise rejection word storage unit that stores rejection vocabularies corresponding to surrounding speech, machine noise, and environmental noise, and a rejection word passed from the rejection word removal unit It is characterized by having a rejection word determination unit that stores a rejection word in the noise rejection word storage unit if it is included in the rejection word database, and an utterance rejection word storage unit if it is not included in the rejection word database. The method comprises the steps of: removing a rejected word that matches; removing a rejected vocabulary that matches a similar vocabulary of the recognized vocabulary; and removing a rejected vocabulary using a plurality of recognized vocabulary pronunciation data.

[0011]

Embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram of a first embodiment of a noise rejection apparatus for speech recognition according to the present invention.

A noise rejection apparatus for speech recognition according to a first embodiment comprises a speech input unit 1 for receiving a human speech utterance and outputting a speech input signal, and a speech recognition result based on the inputted speech input signal. A voice recognition unit 2 that outputs the text in a text format or the like;
The voice output unit 3 outputs the obtained voice recognition result to an external output device or another program.

The recognition dictionary 4 includes a recognition word storage unit 41 for storing recognition words and a rejection word storage unit 42 for storing vocabularies for rejecting invalid speech utterances and noises. Is done.

Further, the recognition dictionary unit 4 removes from the rejected word database 5 the rejected word database 5 that comprehensively stores the rejected words and the vocabulary that matches the recognized vocabulary passed from the recognized word storage unit 41. And a rejected word removing unit 6 for storing the vocabulary in the rejected word storage unit 42.

The voice input by the voice input unit 1 is
Based on the recognition vocabulary stored in the recognition dictionary unit 4, the speech recognition unit 2
And the result is output to the speech recognition result output unit 3. The recognition dictionary unit 4 includes a recognized word storage unit 41 and a rejected word storage unit 42, and can discriminate whether the result obtained by the speech recognition result output unit 3 is a recognized word or a rejected word. As a result, when the speech recognition result is included in the rejected word storage unit 42, it is possible to take measures such as rejection or returning an erroneous recognition reaction.

Next, a method of generating a rejected vocabulary stored in the rejected word storage section 42 will be described.

[0018] The speech recognition vocabulary in FIG. 6 ( "Good morning", "Hello", "Good evening") is "◆" are mapped to the space, the voice recognition space surrounded by a dotted line of the circle surrounding the speech recognition engine correct It shall be interpreted as a recognition result. It should be noted that, when there is overlap in the circle, such as "Hello", "Good evening" in FIG. 2, distance voice recognition vocabulary of the closer of becomes the recognition result. The radius of this dotted circle is the threshold for returning the speech recognition result. The lower the rejection judgment threshold, the larger the radius. On the other hand, the space surrounded by the solid line frame inside the circle means the space of the uttered voice to be correctly recognized.

Since an unspecified speaker and uttered voices from various distances from a short distance utterance to a long distance utterance are included in objects to be correctly recognized, the space of the solid line frame is inevitably increased. In order to absorb the space of the solid line frame, it is necessary to lower the threshold value and increase the radius, and as a result, unnecessary voices easily enter the inside of the dotted circle, so that erroneous recognition increases.

It is to be noted that the environmental noise and the uttered voice can be distinguished, the environmental noise can be absorbed by a relatively short syllable number, and an image diagram of the speech recognition space after the rejection dictionary is constructed is shown in FIG.

It is an image of the rejection dictionary construction that the space other than around the speech recognition vocabulary is covered by the speech recognition range of the rejection word due to the dissemination of the rejection word. And
The speech recognition space is divided into an area where environmental noise is recognized and another area where human speech is mainly recognized.
A noise rejection dictionary consisting of noise rejection words covering the environmental noise area (x in the figure) and a speech rejection dictionary consisting of speech rejection words covering other areas (● in the figure) can be constructed separately. It is.

FIG. 8 shows an example of the rejection word database 5. It is shown as an example "Hello" as the speech recognition vocabulary.

The rejected word database 5 is in the form of a syllable net in which syllables of one to several characters are connected. In FIG. 6, four characters are used as an example.

Regarding the number of characters of the syllable net in the rejected word database, if it is too short, sufficient rejection performance cannot be obtained, and if it is too long, words registered in the recognition dictionary 4 of the recognized word storage unit 41 and the rejected word storage unit 42 There is a problem that the number is enormous and it takes time for voice recognition. For Japanese, about four characters are considered appropriate.

On the other hand, since there is an improvement in the storage capacity and the calculation speed of the computer, it is not particularly necessary to stick to four characters, but it is effective to create a rejection dictionary.

FIG. 9 is a flowchart showing the operation of the rejected word removing unit 6.

First, the rejected word removing unit 6 copies all rejected words contained in the rejected word database 5 (S1). Next, all the recognized words passed from the recognized word storage unit 41 are removed from the rejected words (S2), and the remaining rejected words are output (S3).

The step of removing all recognized words passed from the recognized word storage section 41 of FIG. 9 from rejected words (S2) will be described in detail.

The rejection dictionary removes rejected words matching the speech recognition vocabulary from the rejection word database including all the character strings composed of four characters. It can be configured by removing rejected vocabulary using speech recognition vocabulary utterance data of multiple speakers.

[0030] similar vocabulary in here in the course 2, "Hello" first four characters of "Today,""child" of = "ko", "Hon'nichi" from the fact that k and easy to consonants is erroneous recognition of h of Is removed.

In step 3, the correct speech is recognized by the speech recognition dictionary, and when the rejected word is erroneously recognized, the process of reconstructing the speech dictionary from which the rejected word has been removed and re-recognizing it is repeated.

[0032] In FIG. 6, is an example of "Hello" is recognized right after reject the word "Konichi,""Konii,""Hon'ichi" is erroneously recognized.

Step 3 may be omitted if a certain degree of erroneous recognition is allowed.

Although FIG. 6 has described the speech vocabulary rejection dictionary, a method similar to the speech vocabulary rejection dictionary can be used for environmental noise.

At this time, the character string may be smaller than the character string used in the speech vocabulary.

It should be noted that there is no problem even if the number of characters of the environmental noise is smaller than that of the database of the speech vocabulary. Further, in addition to the speech sound, environmental noise, for example, the sound of knocking on a door and the sound of a fire alarm are put in the recognition dictionary as recognition words. Next, referring to FIG. 2, a second embodiment of the present invention will be described. The embodiment will be described.

The noise rejection apparatus for speech recognition according to the second embodiment has the same configuration as that of the first embodiment shown in FIG. A similar word generation unit 7 that generates a word and outputs it to the rejected word removal unit 6 is configured. The similar word generation unit 7 generates a similar word of the recognized word including the received recognized word itself, and outputs it to the rejected word removing unit 6.

In speech recognition, it is rare that a user's utterance is recognized as acoustic and the result is as expected.
Therefore, when removing a rejected word based on a recognized word, it is necessary to generate and remove a similar word of the recognized word.

Various methods can be considered for generating similar words. Here, two types of similar word generation methods will be described.

One is acoustic similar word generation, and the other is similar word generation using partial matching. First, in the acoustic synonym generation, pairs that are likely to cause misrecognition at the phoneme level are picked up in advance, and a syllable having the same vowel as a consonant (these are called similar syllables) is recognized. Generates synonyms for. For example, '
If the phonemes of K '(ka line) and' H '(c line) are registered as similar phonemes and the recognized word has a syllable of ka line, the syllable replaced with a similar syllable is registered as a similar word I do.

FIG. 10 shows a flowchart of an algorithm for generating a similar word for one recognized word. First, a recognized word is registered as a similar word (S4). Cut out the first syllable of the recognized word (S5), and if there is a similar syllable for that syllable (S6), add a similar word that replaces the corresponding similar syllable for all registered similar words (S6) S7). In the processing of S6 and S7, syllables are cut out (S8) for all syllables of the recognized word (S9).

Using this algorithm, 'K' (ka row) and '
When the phoneme of H '(ha line) is registered as a similar phoneme, and "Kenkai (view)" is input to the similar word generator as a recognition word, "Kenkai", "Henkai", "Kenhai"
The four words "Henhai" are generated as similar words.

Next, a method of generating a similar word using a partial match will be described. Synonyms that use partial matching are
This is one in which a fixed partial length of continuous partial syllables is fixed for the recognition word, and the remaining syllables are changed to arbitrary syllables. FIG.
1 is fixed to one character for the recognition vocabulary
This shows a synonym using a two-character fixed partial match.

When the number of syllables of the recognized word is longer than the maximum number of syllables in the rejected word database, a character string having the maximum number of syllables is cut out from the first character of the recognized word and used.

The shorter the fixed partial syllable length, the more similar words are generated. As a result, it is possible to remove the rejected vocabulary, but the rejection performance deteriorates. Conversely, if the partial syllable length is increased, similar words are reduced and rejection performance is too strong. It is considered that the fixed partial syllable length should be determined according to the number of recognized words.

It should be noted that acoustic similar words and similar words by partial matching can be generated in combination.

Next, a third embodiment of the present invention will be described with reference to FIG. The noise rejection apparatus for speech recognition according to the third embodiment has the same configuration as that of the second embodiment shown in FIG. Rejection word determination that acquires a recognition result when a correct answer voice is input from the voice input unit 8 and the speech recognition result output unit 3 and outputs the recognition result to the rejection word removal unit when the recognition result is a rejection word It comprises a part 9.

In the third embodiment, rejected words that cannot be dealt with by rejected word removal based on recognized words are removed based on the user's voice. The correct voice input unit 8 reproduces a voice in which a correct answer has been obtained by voice recognition of only the recognized word (this is called a correct voice) in advance. The speech recognition result of the correct speech is passed from the speech recognition result output unit 3 to the rejected word determination unit 9.
If the received speech recognition result is not correct and is incorrectly recognized as a vocabulary included in the rejected word storage unit 42, the speech recognition result is passed to the rejected word removing unit 6. The rejected word removing unit 6 stores in the rejected word storage unit 42 what is obtained by removing the recognized word, its similar words, and the speech recognition result from the rejected word database 5 according to the flowchart shown in FIG. Since it is actually necessary to continuously play the correct speech and remove the vocabulary erroneously recognized as a rejected word, the rejected word removal determination unit 9 stores all rejected words that have been removed in the past, It is necessary to transfer to the rejected word removing unit 6 including the rejected word to be removed.

The correct speech may be a specific speaker's correct speech if it is speech recognition by a specific speaker. It is desirable to use audio.

Next, a fourth embodiment of the present invention will be described with reference to FIG.

The noise rejection apparatus for speech recognition according to the fourth embodiment differs from the configuration diagram of the third embodiment shown in FIG. The rejected word is output to the rejected word removing unit 6 after generating the similar word of the rejected word to be removed.

By removing in advance the vocabulary similar to the vocabulary erroneously recognized as the rejected word, the trouble of removing the rejected word by the correct speech input can be reduced, but the rejected word which does not hinder the correct speech input is also removed. It may weaken rejection performance.

The method according to the fourth embodiment is similar to the method according to the third embodiment.
It is considered to be effective when it takes an enormous amount of time to remove rejected words using the correct speech.

Next, a fifth embodiment of the present invention will be described with reference to FIG.

The noise rejection apparatus for speech recognition according to the fifth embodiment is different from the configuration according to the third or fourth embodiment in that
The rejection word storage unit 42 includes an utterance rejection word storage unit 421 and a noise rejection word storage unit 422, and includes a noise rejection word database 10 that stores rejection vocabulary corresponding to surrounding speech, machine noise, and environmental noise, and a rejection word removal. If the rejected word passed from the unit 6 is included in the noise rejection word database 10, the rejection word determination unit 11 stores the rejection word in the noise rejection word storage unit 421. It is comprised including.

The noise rejection word database 10 is constructed by inputting noise speech recorded in advance to a speech recognition dictionary for all vocabularies in the rejection word database, collecting and storing the recognition results. It is desirable that the input noise voice includes various noise sources in consideration of an environment for performing voice recognition.

It should be noted that noisy speech does not include speech when the user speaks other than the recognized word. The rejected word determination unit 11 checks whether or not the rejected vocabulary passed from the rejected word removal unit 6 is included in the noise rejected word database 10.
If not included, it is stored in the utterance rejection word storage unit 421.

By rejecting such rejected words, most of the noise speech is recognized by the vocabulary included in the noise rejected word storage unit, and erroneous utterances other than the recognized words are recognized by the vocabulary included in the utterance rejected word storage unit. It will be easier.

[0059]

According to the present invention, a vocabulary affecting a recognition word required for speech recognition is removed from a huge rejection word database using a syllable net in advance, and the generated rejection word is incorporated into a speech recognition dictionary. Thus, it is possible to reject speech other than noise and the recognized word, and return a correct recognition result only when the recognized word is correctly uttered.

Further, by removing rejected words based on the result of recognizing the correct speech, not only a speech recognition interface premised on clearly speaking at a short distance, but also noise rejection adapted to the use form of speech recognition. A dictionary can be generated.

Further, by dividing the rejected words into noise rejected words and speech rejected words, it is possible to distinguish between noise speech and erroneous utterances other than the recognized words. As a result, it is possible to construct an interface that disregards the noise voice and teaches the user about the erroneous utterance.

[Brief description of the drawings]

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

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

FIG. 3 is a configuration diagram of a third embodiment of the present invention.

FIG. 4 is a configuration diagram of a fourth embodiment of the present invention.

FIG. 5 is a configuration diagram of a fifth embodiment of the present invention.

FIG. 6 is a diagram showing an image of a speech recognition space.

FIG. 7 is a diagram showing a speech recognition space image after the rejection dictionary is constructed.

FIG. 8 is a diagram illustrating an example of a noise rejection database.

FIG. 9 is a flowchart showing an operation of a rejected word removing unit.

[Explanation of symbols]

 Reference Signs List 1 voice input unit 2 voice recognition unit 3 voice recognition result output unit 4 recognition dictionary unit 41 recognized word storage unit 42 rejected word storage unit 421 uttered rejected word storage unit 422 noise rejected word storage unit 5 rejected word database 6 rejected word determination unit 7 Similar word generator 8 Correct speech input unit 9 Rejected word determination unit 10 Noise rejection word database 11 Rejected word determination unit

Claims (8)

[Claims] 1. A voice input unit that inputs voice and outputs a voice input signal, a voice recognition unit that outputs a voice recognition result based on the voice input signal, and a voice output unit that outputs the voice recognition result. A speech recognition system having a recognition dictionary used for speech recognition in the speech recognition means, wherein the recognition dictionary is for recognizing a recognition word storage for storing a recognition word and invalid speech utterance or noise. And a rejection word database that stores the rejection words comprehensively, and a vocabulary matching the recognition vocabulary passed from the recognition word storage unit is removed from the rejection word database. An apparatus for creating a recognition dictionary, comprising: a rejection word removing unit that stores remaining vocabulary in the rejection word storage unit. 2. The recognition dictionary according to claim 1, further comprising: a synonym generation unit that generates a synonym of the recognition word passed from the recognition word storage unit and outputs the synonym to the rejection word removal unit. Creating device. 3. A correct speech input means for inputting a correct speech which uttered the recognized word correctly to the speech recognition means, and acquiring the recognition result when the correct speech is input from the speech recognition result output means, 3. The recognition dictionary creating apparatus according to claim 2, further comprising: a rejection word determining unit that outputs the recognition result to the rejection word removing unit when the recognition result is the rejection word. 4. The rejected word determining means outputs the recognition result to a similar word generating means instead of a rejected word removing means, and outputs the similar word to the rejected word removing means. Item 3. The recognition dictionary creation device according to Item 3. 5. A rejection word database, wherein the rejection word storage means comprises speech rejection word storage means and noise rejection word storage means, and stores the rejection vocabulary corresponding to surrounding speech, machine noise, and environmental noise. If the rejected word passed from the rejection word removing unit is included in the noise rejection word database, the rejection word is stored in the noise rejection word storage unit, and if not, the rejection word is stored in the utterance rejection word storage unit. The apparatus according to claim 3, further comprising a rejection word determination unit. 6. A recognition dictionary used for voice recognition, wherein the recognition dictionary stores a recognition word storing a recognition word and a rejection word storing a vocabulary for rejecting invalid speech utterance or noise. The rejection word storage unit includes a recognition word storage unit for storing a recognition word, an utterance rejection word storage unit for storing a vocabulary for rejecting invalid speech utterance and noise, and an environmental noise rejection word storage area. A word dictionary characterized by the following. 7. A method for generating a rejection dictionary comprising the steps of: removing a rejected word matching a recognized vocabulary from a rejected word database; and removing a rejected vocabulary matching a similar vocabulary of the recognized vocabulary. 8. A step of removing a rejected word matching a recognized vocabulary from a rejected word database, a step of removing a rejected vocabulary matching a similar vocabulary of the recognized vocabulary, a rejected vocabulary using a plurality of recognized vocabulary pronunciation data. And generating a rejection dictionary.