JPS6039522A - Word voice recognizing method - Google Patents

Abstract

PURPOSE:To maintain a high recognition rate by correcting a value of tolerance weight in the direction for reducing a difference between the first rank of a weighted similarity degree and a weighted similarity degree in a dictionary item of an input word, whenever an erroneous recognition is generated, and using its value in the next time and thereafter. CONSTITUTION:When a word A is at the first rank of a similarity degree, for instance, if a value of tolerance weight applied to a similarity degree of a word B is an (n) point, and a result of work recognition is A against an input of B, the previous value of tolerance weight is corrected automatically to (n)+(m) points. Also, in case when a correct recognition result is obtained against an input of A and B, a value of tolerance weight remains as it is. On the other hand, in case when the result of word recognition becomes B against the input of A, on the contrary, the value of tolerance weight is corrected to (n)-(m). By continuing it, the value of tolerance weight is converged to a value by which a recognition rate as a whole becomes the highest.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a word speech recognition method for recognizing words by comparing input speech with a word dictionary in which phonemes are expressed.

The configuration of a conventional example and its problems A conventional word speech recognition method will be explained with reference to FIG. As shown in the figure, the input speech is first analyzed, the features of the input word speech are extracted, and the phonemes that make 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 the degree of similarity between the two phoneme sequences is calculated as follows:
It is calculated by calculating the recognition probability for each phoneme using a contusion matrix (CM) between phonemes, and then for each combination of each dictionary item and the dictionary item with the highest similarity. A predetermined likelihood of 1 is added to or subtracted from the similarity for each dictionary item, and the dictionary item with the maximum weighted similarity 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.

Table 1, Table 2, Table 3, and Table 3 show a portion of the conflation matrix. In Table 3, the vertical lines indicate phonemes in the word dictionary, and the horizontal lines indicate recognized phonemes. Furthermore, the numbers in Table 3 indicate the probability of what kind of phoneme each phoneme in the word dictionary will be recognized as. For example, in Table 3, the probability that l in the word dictionary will be recognized as ■ is 75%, the probability that it will be recognized as U is 5chi, the probability that it will be recognized as A is oq6, the probability that it will be omitted is 8%, etc. It shows.

In the above conventional example, if the recognized phoneme sequence of word A has a unique reverse tendency for phoneme recognition errors, and such recognized phoneme sequence has a higher degree of similarity with the dictionary phoneme sequence of word A than word B, then the word The purpose of using the likelihood weight is to adjust the similarity value in order to prevent A's input speech from being incorrectly recognized by B in a concentrated manner.

The value of this likelihood weight was set to a value that maximized the word recognition rate by performing recognition experiments on a large number of voice data in advance. If the phoneme recognition tendency for each word does not always change, the conventional method can be used, but if the phoneme recognition tendency changes due to the influence of the transmission path, ambient noise, etc., the likelihood weight value will not be the optimal value. There was a problem that the word recognition rate decreased.

OBJECTS OF THE INVENTION It is an object of the present invention to provide a word speech recognition method that solves the problems of the conventional example and can maintain a high word recognition rate even if the environment changes.

Structure of the Invention In order to achieve the above object, the word speech recognition method of the present invention changes the likelihood weight value used to calculate the one-time similarity between the input speech and the dictionary phoneme sequence to the actual word speech. Every time a misrecognition occurs during recognition, the difference between the weighted similarity at that time and the weighted similarity in the dictionary entry of the input word is corrected to reduce the difference, and the corrected value is used for the next word speech. It is characterized by being used for recognition.

DESCRIPTION OF THE EMBODIMENTS The word speech recognition method in this embodiment is an improvement on the conventional example, and the outline of the recognition algorithm is shown in the diagram as in the conventional example. Performs phoneme recognition of the input voice, calculates the degree of similarity between this recognized phoneme sequence and the dictionary phoneme sequence of each dictionary item in the word dictionary, and calculates weighted similarity by adding or subtracting a likelihood weight to this degree of similarity. However, it is similar to the conventional example in that the dictionary entry with the maximum value is selected as the recognized word, but the value of the likelihood weight used is not a fixed value, but is modified to an optimal value according to changes in phoneme recognition trends. This is different from the conventional example. The revised method will be described in detail. When word A has the highest similarity, the value of the likelihood weight added to the similarity of word B is n points, and if the word recognition result is A for the input of B, then the value of the previous likelihood weight is automatically corrected to n 10 m points, and in subsequent recognition this n
The 10m point will be used. Further, if correct recognition results are obtained for inputs A and H, the value of the likelihood weight remains unchanged. Further, when the word recognition result is B for input A, the value of the likelihood weight is corrected to n - m, contrary to the above. By continuing this process, the value of the likelihood weight converges to a value where the probability of misrecognizing A and B is equal, that is, the overall recognition rate is the highest. Note that even if the initial value is 0 point or a value obtained experimentally in advance, if the correction is repeated, the value will converge to the same value.

In this embodiment, even if the phoneme recognition tendency changes due to the influence of the transmission path, ambient noise, etc., a corrected value of the likelihood weight can be obtained while performing word speech recognition, and this can be used for the next word speech recognition. This has the advantage that a high recognition rate can be maintained.

Note that the present invention is applicable not only to the case where a recognized phoneme sequence is uniquely determined, but also when a recognized phoneme sequence in a lattice format is used, and furthermore, in a word speech recognition algorithm, the obvious form of the recognized phoneme sequence? A similar effect can also be obtained when applying the method to directly match the input speech analysis result with the dictionary phoneme sequence of each item in a word dictionary with phoneme notation. can.

Effects of the Invention According to the present invention, there is an advantage that a high word recognition rate can be maintained even if the environment such as the transmission path and ambient noise changes.

[Brief explanation of the drawing]

The figure is a diagram illustrating an outline of a word speech recognition method for explaining an embodiment of the present invention and a conventional example thereof.

Claims (1)

[Claims] When recognizing words by calculating the similarity between the input speech and the dictionary phoneme series of each dictionary entry in the word dictionary with phoneme notation,
When calculating the similarity for the input speech, a likelihood weight predetermined for each combination of each dictionary item and the dictionary item with the highest similarity is added to the similarity for each dictionary item. In a word speech recognition method, a weighted similarity is calculated by addition, subtraction, multiplication, or division, and the dictionary entry with the highest weighted similarity is used as a recognized word. A word sound characterized in that each time recognition occurs, the difference between the current weighted similarity degree and the weighted similarity degree in a dictionary entry of the input word is automatically corrected in a direction that reduces the difference. Recognition method.