JPS595292A - Word voice recognition method - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a word speech recognition method that first performs phoneme recognition on input speech and then recognizes words by comparing this recognized phoneme sequence with a word dictionary in which phonemes are expressed. The present invention provides a word speech recognition method that achieves a high word recognition rate.

A conventional word recognition method will be explained with reference to FIG.

As shown in FIG. 1, input speech is first analyzed, features of the input word speech are extracted, and phonemes making up the input word speech are recognized. This recognized phoneme sequence is compared with the dictionary phoneme sequence of each dictionary entry in the word dictionary, and 2
The likelihood between two phoneme sequences is calculated by finding the recognition probability for each phoneme using a confusion matrix between phonemes (ConfusionMatriX + hereinafter abbreviated as C, M). The dictionary entry with the maximum likelihood is selected as a recognized word.

Table 1 shows an example of a word dictionary used in the word speech recognition method, and each word is written according to the phoneme notation shown in Table 2. FIG. 2 shows a part of the above-mentioned C and M. In FIG. 2, the vertical lines indicate phonemes in the word dictionary, and the horizontal lines indicate recognized phonemes.

Further, the numbers in FIG. 2 indicate the probability of what kind of phoneme each phoneme in the word dictionary will be recognized as. For example, in Figure 2, the probability that the phoneme word in the word dictionary will be recognized as "work" is 75%, the probability that it will be recognized as "UK" is 5%, the probability that it will be recognized as "A" is 0%, and the probability that it will be omitted is 8%. ...etc.

Table 1, Table 2. In the case of a word that has a phoneme with a low phoneme recognition rate, the likelihood of the recognized phoneme sequence of the input speech of that word and the dictionary phoneme sequence is phoneme 1.
Even if the identification result is good, the difference in likelihood between the dictionary entries of other words and the dictionary phoneme series tends to be small.

For such words, if there is an error in the phoneme recognition of the input voice, the correct dictionary entry (for the input voice of a certain word,
(The dictionary entry for that word is called the correct dictionary entry.) A situation easily arises in which the likelihood between phoneme sequences is higher for other dictionary entries than for other dictionary entries, and this is the case when conventional word recognition methods result in incorrect word recognition. This caused a decline in the word recognition rate.

The present invention is intended to significantly improve the drawbacks of the above-mentioned conventional examples, and one embodiment of the present invention will be described below.

In this embodiment, as in the conventional example shown in FIG. seek. The process up to this point is the same as the conventional example, but a weighted likelihood value is calculated by adding a likelihood weight value predetermined for each dictionary item to this likelihood, and this weighted likelihood value is The word with the maximum number of dictionary entries is recognized. That is, words are recognized using weighted likelihood values rather than the likelihood itself.

Next, a method of determining the likelihood weight value in this embodiment will be explained. First, in a conventional word speech recognition method, word recognition is performed using speech data.

Then, using only the speech data for which the word recognition result was correct, the average likelihood between the recognized phoneme sequence for each word and the dictionary phoneme sequence of the correct dictionary item is calculated.

The average likelihood value for each word obtained here represents the difficulty level of phoneme recognition for each word. As described in the explanation of the conventional example, for words with a low average likelihood value, there is a strong possibility that other dictionary items have a higher likelihood than the correct dictionary item. Note that the reason why we used only the speech data with correct word recognition results to calculate the average likelihood value is because speech data with incorrect word recognition is often defective in itself.
This is to eliminate such data. By the way, the likelihood between phoneme sequences is obtained by adding the logarithmic recognition probabilities of each phoneme constituting the phoneme sequence, and normalizing the result to give a perfect score of 900 regardless of the number of phonemes. In this example, the likelihood weight value for each word dictionary item was calculated using the following equation.

Next, an example of word recognition results according to this embodiment will be shown in comparison with a conventional example. Word "Matsubara" (dictionary phoneme series MAC
UBARA) The phoneme recognition results of Chiru's voice data are A P
It was UB AWA. At this time, the likelihood with the correct dictionary item was 788 points, and the likelihood with the dictionary item KAKOG*AWA was 790 points. In this case, the word recognition result according to the conventional example was "kakogawa" and the word was incorrectly recognized. According to this embodiment, the predetermined likelihood weight value is 7 points for MACUBARA and 2 points for KAKOG*AWA, so the likelihood value determined in advance is 7 points for MACUBARA.
The score is 95 points, and the score is 792 points for KAKOG Sai AWA, and the word is correctly recognized.

As described above, the word speech recognition method of the present invention has the advantage of being able to improve the word recognition rate for words for which the word recognition rate was low using conventional methods.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram showing a word speech recognition method in a conventional example and an embodiment of the present invention, and FIG. 2 is a diagram showing a portion of C and M in the conventional example and an embodiment of the present invention. .

Claims (1)

[Claims] When performing phoneme recognition on input speech to obtain a recognized phoneme sequence, and calculating the likelihood between this recognized phoneme sequence and the dictionary phoneme sequence of each dictionary entry in a word dictionary with phoneme notation, when recognizing a word, A likelihood weight value is determined in advance for each dictionary item, and a weighted likelihood value is calculated by adding or multiplying this likelihood weight value to the likelihood between the phoneme sequences. A word speech recognition method characterized in that a dictionary item with a maximum value is recognized as a word.