JPH09244692A - Uttered word certifying method and device executing the same method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To correctly judge whether an uttered word is of a correct uttered content or not by recognizing it as it is when an acoustically similar word is uttered and comparing it with a recognition objective word directly. SOLUTION: A word list generating part 4 generates a word list to be used for a word voice recognition in addition to pseudo words generated in a pseudo word generating part 2 and KANA(Japanese syllabary) notations preliminarily set in a word setting part 1 to transmit them to a voice recognition part 6. The voice recognition part 6 performs a voice recognizing processing based on the input voice data inputted from a voice input part 5 and the word list generated in the word list generating part 4 and outputs the KANA notation of the word given with the highest likelihood among words registered in the word list as the recognition result. A word certifying part 7 compares the KANA notain being the recognition result of the input voice to be outputted by the voice recognition part 6 with KANA notations of the words inputted to the word setting part 1 and when they coincide, the part 7 judges that an objective word is uttered.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a spoken word authentication method and an apparatus for implementing this method, and more particularly to speech recognition using a preset word and a word acoustically similar to the word as an authentication target vocabulary. The present invention relates to a spoken word authentication method for authenticating that a correct utterance is made when a recognition result is a preset word, and an apparatus for implementing this method.

[0002]

2. Description of the Related Art In authenticating whether or not a speaker who is a user correctly utters a preset authentication target word, conventionally, a voice recognition process is performed by using only the set word as a recognition target vocabulary, Whether or not the utterance content is a preset word is determined based on the magnitude of the likelihood obtained as a result of recognition.

Further, a model for general human speech called a garbage model is created by using speech data of all syllables or phonemes of many speakers.
While the speech recognition processing is performed to calculate the likelihood given to the uttered speech data by this model, the likelihood of the input speech data for the syllable or phoneme sequence of the word set in advance is calculated. A spoken word authentication method has also been implemented in which it is assumed that a preset word is uttered when there is a significant difference in comparing degrees.

[0004]

However, the uttered word authentication method for recognizing only the set words as the vocabulary to be recognized is liable to be affected by noise or other environment in which the speech recognition is performed. , There is a possibility that some speakers may get a low likelihood even if they speak a set word, so it is hard to say that it is a spoken word authentication method with a strict meaning.

In the uttered word authentication method using the garbage model, when the voice recognition process is performed in syllable units, the likelihood of the preset syllable string of the word with respect to the input voice data can be calculated. Since no relative comparison is made between this likelihood and the likelihood given by a word having a syllable sequence similar to the set word, a syllable or a syllable that is partially different from the preset word is not compared. Has the possibility of erroneously authenticating a word with a syllable sequence when it is spoken.

The present invention provides a spoken word authentication method that solves the above-mentioned problems and an apparatus that implements this method.

[0007]

[Means for Solving the Problems] Speech recognition is performed using a preset word and a word acoustically similar to the word as a recognition target vocabulary, and a recognition result of an input voice and a preset word are compared. By comparison, a uttered word authentication method was constructed to authenticate that the correct utterance was made when the recognition result was a preset word.

Then, a vocalized word authentication method is constructed in which each syllable of the set word is replaced with an acoustically similar syllable to create a similar word. We also constructed a spoken word authentication method that refers to a syllable conversion table when replacing each syllable of a set word with an acoustically similar syllable. Furthermore, a vocabulary word authentication method that selects an arbitrary syllable when a similar syllable is selected from the syllable conversion table is constructed.

Then, when similar syllables are selected from the syllable conversion table, a desired number of syllables are sequentially selected by referring to a statistic representing the ease of replacement between syllables.
We constructed a spoken word authentication method that creates multiple words. Here, a pseudo word generation unit 2 that includes a word setting unit 1 in which a word to be authenticated is preset and that generates a pseudo word having a similar syllable sequence based on the word input from the word setting unit 1 is used. The pseudo word generated by the pseudo word generation unit 2 is provided with the syllable conversion table 3 that stores the syllables that are acoustically similar and that easily cause mutual recognition errors and supply the syllables to the pseudo word generation unit 2. A word list generation unit 4 that creates a word list used for word speech recognition by combining the word and a word preset in the word setting unit 1 is input, and the uttered voice is input and converted into voice data. A voice input unit 5 is provided, and voice recognition processing is performed based on the input voice data input from the voice input unit 5 and the word list created by the word list generation unit 4, and the words registered in the word list are processed. The most of The speech recognition unit 6 that outputs a word given a high likelihood as a recognition result is provided, and the recognition result of the input speech output by the speech recognition unit 6 is compared with a word preset in the word setting unit 1. A voiced word authentication device including a word authentication unit 7 that determines that a target word is uttered when the two match.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described with reference to the example of FIG. 1 is a word setting unit, 2 is a pseudo word generation unit, 3 is a syllable conversion table, 4 is a word list generation unit, 5 is a voice input unit, 6 is a voice recognition unit, and 7 is a word authentication unit. In the word setting unit 1, kana notation of a word to be authenticated is input. The input kana notation word is input to the pseudo word generation unit 2. The pseudo word generation unit 2 generates a desired number of pseudo words having similar syllable strings based on the input kana notation word and sends the pseudo word to the word list generation unit 4 as a registered word.

Here, a method of generating a pseudo word will be described. FIG. 2 shows an example of the syllable conversion table 3. The syllable conversion table 3 of FIG. 2 stores, for each class, syllables that are acoustically similar and easily cause mutual recognition errors. For example, the syllable "ha" stored in the class 1 of FIG. 2 is erroneously recognized as "a" due to the lack of consonants, or erroneously recognized as "sa", "ta" or other syllables due to the replacement of consonants. Is likely to be. Further, in the example shown in FIG. 2, syllables that include the sound of the same vowel are stored in different classes.

A given kana expression word is decomposed into syllables, and syllable conversion tables are provided for each syllable randomly based on the input syllables and syllables whose syllable sequences are similar to each other, or based on a previously calculated syllable appearance frequency. It is selected from 3. The selected syllables are sequentially connected, and one pseudo word is generated.
The pseudo word generation unit 2 repeats the above processing to generate a desired number of pseudo words. Further, the pseudo word generation unit 2
When the above process is performed multiple times, whether or not a pseudo word having the same syllable string as the already generated pseudo word is generated so that all generated pseudo words have different syllable strings. To judge. If a pseudo word having the same syllable string has already been generated, the newly generated pseudo word is rejected. On the other hand, if a pseudo word having the same syllable string has not been generated yet, the newly generated pseudo word is sent to the word list generation unit 4 as a registered word.

FIG. 3 shows an example of pseudo word generation. In this example, “KAIEDA Shoichi” is input to the word setting unit 1 as a kana notation word. The input kana notation word “Kaieda Shoichi” is input to the pseudo word generation unit 2. In the generation of the pseudo word, the pseudo word generation unit 2 refers to the syllable conversion table 3 and selects a similar syllable for each given syllable. Here, when similar syllables are selected from the syllable conversion table 3, a desired number of syllables are sequentially selected with reference to a statistic representing the ease of replacement between syllables, and a plurality of words are created. Pseudowords are generated that lead to less. In the example shown in FIG. 3, "ta" is selected for "ka", "ri" is selected for "i", "re" is selected for "e", and so on. After that, these were connected in sequence to generate the pseudo word "Tare Leanyo Fumiji". The pseudo word generator 2
The pseudo word “Tare Lean Fumiji” thus generated is sent to the word list generation unit 4 as a registered word.

The word list generator 4 is a pseudo word generator 2.
The word list used for word voice recognition is created by combining the pseudo-word generated in step 1 and the kana notation of the word to be authenticated, which is preset in the word setting unit 1, and the voice recognition unit 6
Send to FIG. 4 shows a word list generation example in the word list generation unit 4. Although the pseudo word generator 2 can generate an arbitrary number of pseudo words, the example shown in FIG. 4 is an example in which ten pseudo words are generated. All the generated pseudo-words are combined with the kana notation preset in the word setting unit 1, and the word list shown on the right side of FIG. 4 is created. The word list created by the word list generation unit 4 is sent to the voice recognition unit 6 and set as a recognition target vocabulary.

The voice recognition unit 6 performs a voice recognition process based on the input voice data input from the voice input unit 5 and the word list created by the word list generation unit 4.
The kana notation of the word given the highest likelihood among the words registered in the word list is output as the recognition result.
The word authentication unit 7 compares the kana notation, which is the recognition result of the input voice output by the voice recognition unit 6, with the kana notation of the word input to the word setting unit 1, and when both kana notations match. Determines that the target word has been uttered. On the other hand, when the kana notation does not match, it is determined that the wrong word is uttered.

[0016]

As described above, according to the present invention, when an acoustically similar word is uttered, it is recognized as it is and directly compared with the word to be authenticated. Can be accurately determined. The difficulty of speech recognition is approximately proportional to the number of recognition target words, but according to the present invention, the number of recognition target words can be freely set by adjusting the number of pseudo words to be generated. Therefore, the difficulty of voice recognition can be easily changed by this setting, and the strictness of uttered word authentication can be freely set.

Further, by changing the definition of the similar syllable described in the syllable conversion table, it is possible to easily change the characteristics of the similar word to be generated.

[Brief description of drawings]

FIG. 1 illustrates an embodiment.

FIG. 2 is a diagram showing a syllable conversion table.

FIG. 3 is a diagram showing an example of pseudo word generation.

FIG. 4 is a diagram showing an example of word list generation.

[Explanation of symbols]

 1 word setting unit 2 pseudo word generation unit 3 syllable conversion table 4 word list generation unit 5 voice input unit 6 voice recognition unit 7 word authentication unit

Claims (6)

[Claims] 1. A speech recognition is performed by using a preset word and a word acoustically similar to the preset word as a recognition target vocabulary, and recognition is performed by comparing a recognition result of input speech with a preset word. A uttered word authentication method, characterized in that if the result is a preset word, it is authenticated that the correct utterance has been made. 2. The spoken word authentication method according to claim 1, wherein a similar word is created by replacing each syllable of the set word with an acoustically similar syllable. Authentication method. 3. The spoken word authentication method according to claim 2, wherein a syllable conversion table is referred to when replacing each syllable of the set word with a syllable that is acoustically similar. Authentication method. 4. The spoken word authentication method according to claim 3, wherein an arbitrary syllable is selected when a similar syllable is selected from the syllable conversion table. 5. The voicing word authentication method according to claim 3, wherein when a similar syllable is selected from the syllable conversion table, a desired number of syllables are sequentially referred to by referring to a statistic representing the ease of replacement between syllables. A spoken word authentication method characterized by selecting only syllables and creating multiple words. 6. A word setting unit in which a word to be authenticated is preset, and a pseudo word generation unit that generates a pseudo word having a similar syllable sequence based on the word input from the word setting unit. However, it is equipped with a syllable conversion table that stores syllables that are acoustically similar and are prone to mutual recognition errors and supplies them to the pseudo word generation unit. The pseudo word generated by the pseudo word generation unit and the word setting unit It is equipped with a word list generation unit that creates a word list used for word speech recognition together with the preset words, and a voice input unit that inputs the uttered voice and converts it into voice data. , The voice recognition process is performed based on the input voice data input from the voice input unit and the word list created by the word list generation unit, and the highest likelihood of the words registered in the word list is given. When a speech recognition unit that outputs the recognized word as a recognition result is provided, the recognition result of the input speech output by the speech recognition unit is compared with the word preset in the word setting unit, and both match An uttered word authentication device comprising: a word authentication unit that determines that the target word is uttered.