JPH0827639B2 - Voice recognition device - Google Patents

Description

Description: FIELD OF THE INVENTION The present invention relates to an improvement of a voice recognition device for recognizing and outputting a voice uttered in a unit of word / phrase or in a plurality of phrases.

[Background of the Invention]

As described in Japanese Patent Laid-Open No. 59-116837, a conventional device recognizes a voice uttered in a phrase unit, and a condition including an independent word length and frequency other than the certainty of a voice recognition result from a plurality of candidate sequences. Then, the output order of the recognition candidates is evaluated and changed. However, in the device, the recognition performance may be deteriorated when a bunsetsu containing an independent word having a short word length is input, and consideration is not given to the recognition performance when the independent word length is affected. It was

[Object of the Invention]

It is an object of the present invention to provide a voice recognition device that correctly recognizes a voice uttered in units of words, phrases and plural phrases.

[Outline of Invention]

In order to achieve the above object, the present invention has the following configurations.

Recognition means for recognizing spoken voice, candidate series creation means for creating a plurality of candidate series from the speech recognized by the recognition means, independent word dictionary in which information on independent words is registered, and information on adjunct words In the speech recognition device having an adjunct word dictionary that has registered, and an independent word dictionary and an adjunct word dictionary, and a selection means that selects a correct series from the plurality of candidate series, the selection means is The feature is that the correct sequence is selected based on the number of independent words, the frequency of independent words, and the acoustic similarity.

Example of Invention

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an example of a voice recognition device for explaining the outline of the present invention. The input word / syllable / plural syllable voice is collated with the standard pattern 2 such as phoneme / syllable / word in the acoustic recognition unit 1 and the plural candidate sequences are stored in the recognition result buffer memory 3 together with the acoustic similarity. To do. An example of the contents of the recognition result buffer memory 3 is shown in FIG. The candidate sequence selection unit 4 has an independent word dictionary 6 and an auxiliary word dictionary 7, and the independent word dictionary 6 is shown in FIG.
The independent word and the notation / appearance frequency of the independent word are registered, and the adjunct word and the notation of the adjunct word are registered in the adjunct word dictionary 7 as shown in FIG.

The candidate sequences in the recognition result buffer memory 3 are sent to the candidate sequence selection unit 4 one by one together with the acoustic similarity, and are compared and collated with the character strings of the respective items in the independent word dictionary 6 and the auxiliary word dictionary 7. Decompose into independent words and attached words. The decomposed character string is stored in the sequence solving buffer memory 5 together with the number of independent words, the sequence frequency obtained from the independent word frequency, the unknown word flag, and the acoustic similarity. For example, in the case of the first series “Hitsuyoutearu” in FIG. 2 (a), “Hits”, “U”, and “Al” in the independent word dictionary 6 and “Yo” and “Te” in the auxiliary word dictionary 7 are used. ", And it is disassembled as Hits + Yo + U + Te + Al. At this time, the number of independent words is 3, and the sequence frequency is
(The frequency of "hits") + (the frequency of "c") + (the frequency of "al") divided by the number of independent words, that is, (27 + 34 +
451) / 3 = 171. The unknown word flag has a value of 1 when there is a character string that cannot be matched with the dictionary item. The value of the unknown word flag of the first series “Hitoyo Teal” is 0.
Regarding the 5 sequences in FIG. 2 (a), the number of independent words stored in the sequence solution buffer memory 5, the sequence frequency, the unknown word flag,
The content of the acoustic similarity is shown in FIG.

Next, among the sequences stored in the sequence analysis buffer memory 5, the number of the sequence in which the value of the unknown word flag is 0 is stored in the selected sequence memory 8. In the example, the numbers 1, 2, 3, 4 are stored in the selected sequence memory 8 because the value of the unknown word flag of the sequence 1, 2, 3, 4 is 0. Next, of the series of numbers stored in the selected series memory 8, the number of the series having the smallest number of independent words is stored again in the selected series memory 8. In the example, both the numbers of independent words in the series of 3 and 4 are 2 and the minimum, so that the numbers 3 and 4 are stored in the selection system memory 8. Next, of the series of numbers stored in the selected series memory 8, the number of the series having the maximum acoustic similarity is stored again in the selected series memory 8. In the example, since the acoustic similarity values of the sequences 3 and 4 are both 71, the selected sequence memory 8 has the number 3,
4 is memorized. Next, of the series of numbers stored in the selected series memory 8, the series number having the largest series frequency value is stored in the selected series memory 8 again. In the example, the value of the sequence frequency of the sequence of 3 is 182, and the value of the sequence frequency of the sequence of 4 is 161. Therefore, the number 3 is stored in the selected sequence memory 8.

Finally, the series of numbers stored in the selected series memory 8 is stored in the display buffer memory 9. When performing comparison and collation with each item of the independent word dictionary 6 and the adjunct word dictionary 7, the “notation” of the collated item is stored and finally the number stored in the selected series memory 8 is stored. The "notation" column of the sequence may be stored in the display buffer memory 9.

Then, the contents of the display buffer memory 9 for display are
Display on display 10.

Immediately after the acoustic recognition unit 1, it is possible to narrow down the candidate sequence to a small number by using information such as syllable combinations that can appear in Japanese.

〔The invention's effect〕

As described above, according to the present invention, the correct sequence is selected for the candidate sequence resulting from the speech recognition by using the number of independent words in the sequence, the independent word frequency, and the acoustic similarity. Series are selected and output with high accuracy.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention, and FIG. 2 (a).
Is a diagram showing an output candidate sequence of the acoustic recognition unit when recognizing the phrase "Hitoyodeal", FIG. 7B is a diagram showing an independent word dictionary used in the present invention, and FIG. 7C is used in the present invention. FIG. 3 is a diagram showing an adjunct word dictionary to be used, and FIG. 3 is a diagram showing an example of the number of independent words, the sequence frequency, the unknown word flag, and the acoustic similarity regarding the candidate sequence when "Hitoyodeal" is recognized. 1 ... acoustic recognition unit, 2 ... phoneme / syllable / word standard pattern,
3 ... Recognition result buffer memory, 4 ... Candidate sequence selection unit, 5
... Sequence analysis buffer memory, 6 ... Independent word dictionary, 7 ... Adjunct word dictionary, 8 ... Selected series buffer memory, 9 ... Display memory for display, 10 ... Display.

 ─────────────────────────────────────────────────── ─── Continued Front Page (72) Inventor Masahiro Abe 1-280 Higashi Koigakubo, Kokubunji City, Tokyo Inside Hitachi Central Research Laboratory (72) Inventor Yoshiyuki Takeichi 1-280 Higashi Koigakubo, Kokubunji City, Tokyo Hitachi Ltd. (72) Inventor Hirohide Endo 1-280, Higashikoigakubo, Kokubunji, Tokyo (56) References, Hitachi, Ltd. Central Research Laboratory (56) Reference JP-A-59-180629 (JP, A)

Claims (1)

[Claims] 1. A recognition means for recognizing a spoken voice, a candidate series creation means for creating a plurality of candidate series from the sounds recognized by the recognition means, and an independent word dictionary in which information about an independent word is registered. In a speech recognition device comprising: an adjunct word dictionary in which information about adjunct words is registered; and selecting means for selecting a correct sequence from the plurality of candidate sequences using the independent word dictionary and the adjunct word dictionary. The speech recognition device, wherein the selection means selects a correct sequence from the number of independent words, the independent word frequency, and the acoustic similarity of the plurality of candidate sequences.