JPH0431118B2 - - Google Patents

Description

[Detailed description of the invention]

(Industrial Application Field) The present invention relates to a speech recognition method for recognizing words by comparing input speech with a word dictionary in which phonemes are expressed. (Constitution of Conventional Example and Problems thereof) FIG. 1 is a functional block diagram of an apparatus for executing an example of a conventional speech recognition method and an embodiment of the speech recognition method of the present invention. A conventional example will be explained with reference to FIG. In Fig. 1, 1 is a parameter extraction unit that creates a time series of parameters from input speech, 2 is a word recognition unit that performs segmentation for each phoneme, likelihood calculation, similarity calculation, etc., and 3 is a total number of words to be recognized. This is a word dictionary section that stores a word dictionary in which words are expressed as symbol strings in units of phonemes. For example, the words "Satsuporo", "Tokyo", "Toyonaka", and "Yamakata" are written as "SAQPORO", "TOOKJOO", and "Yamakata" respectively.
It is written as "TOJONAKA", "JAMAGATA", etc. Next, the operation of the above conventional example will be explained. First, the input audio is analyzed by the parameter extraction unit 1 every 10 ms frame, parameters are extracted, and a parameter time series is created. The parameter time series is one in which parameters commonly used in subsequent processing are calculated in advance. Next, the word recognition unit 2 compares the word dictionary unit 3 to determine the degree of similarity for each dictionary item. When measuring this similarity, the phoneme segmentation is performed according to the dictionary phoneme sequence that constitutes the dictionary item. calculate the likelihood, which is a measure of the probability that the segmented speech interval is the one that uttered the phoneme,
The degree of similarity is determined as the average value of the likelihood of each phoneme in the dictionary item, and the dictionary item with the maximum degree of similarity is determined as a recognized word. In this conventional example, the likelihood l _i of the i-th phoneme in the dictionary phoneme sequence is calculated by the following equation. l _i = l _i0 − l _ip1 ... Here, l _i0 is 5000 as a measure of how well the parameters in each frame in the segmented section match the standard pattern of the phoneme prepared in advance. It is calculated on a full point basis. In addition, l _ip1 is used when the time length of the segmented section is less than a predetermined threshold.
l _ip1 = 0, 1 if threshold is exceeded
It becomes a value between 5000 and 5000, and points are deducted from the likelihood. However, the above conventional example had the following drawbacks. This example will be explained with reference to FIG. Figure 2 shows when the input voice is TOJONAKA, the time is set to the right, and the dictionary entry is
FIG. 3 is a diagram showing the correspondence of segmentation results between TOJONAKA and TAMAGATA. In this example, the segmentation and likelihood calculation for the dictionary item TOJONAKA were successfully performed. On the other hand, the segmentation in the case of JAMAGATA included two incorrect correspondences, TOJ--J and A-AGA, but in the likelihood calculation,
Even if the TOJ part of the input is regarded as J, there is no problem in terms of parameters, and since the section segmented as G is an unstable part of the voice transitioning from A to K, there is a small power dip, so Moreover, the parameters matched well with the standard pattern of G,
In addition, since there was no deduction for l _ip1 , a high likelihood was obtained, and the degree of similarity was also large. For this reason, the input voice shown in this example was erroneously recognized as JAMAGATA. Dictionary entry JAMAGAHA shown in this example
In the segmentation in G, the segmented section is 2 frames, and the next A
The segmented section was only one frame. When segmenting a certain phoneme, if the time length of the segment is as short as 1 or 2 frames, the characteristics of that phoneme are weak in utterance, and the transition between that phoneme and the next phoneme is similar to that of the next phoneme. In many cases, the segmentation is performed in intervals of 1 to 2, and therefore, segmentations of short time length are rarely continuous. Therefore, in this conventional example, as in the JAMAGATA example shown in Figure 2, two phonemes are used.
Although the GA performed segmentation using three frames, which is actually impossible, the degree of similarity became large, resulting in words being misrecognized. Furthermore, for the opposite reason to the above, it is rare for segmentations to continue for a very long time, but even in such cases, this conventional example has the disadvantage that the degree of similarity may be large. It was hot. (Objective of the Invention) The present invention eliminates the drawbacks of the above-mentioned conventional examples, and when an improbable or rare segmentation occurs as described above, it adds deduction points to the likelihood, thereby making it possible to recognize words. The aim is to improve the rate. (Structure of the Invention) The present invention collates input speech with each dictionary entry in a word dictionary, and segments the input speech for each phoneme according to the dictionary phoneme series that constitutes each dictionary entry.
The likelihood, which is a measure of the certainty that the segmented speech interval is the one that uttered the phoneme, is calculated, and this likelihood value is used to calculate the degree of similarity between each dictionary item and the input speech. In order to achieve the above objective in recognizing an input word, in calculating the likelihood of a phoneme, one or more previous phonemes are added to the time length of the segmented interval corresponding to that phoneme. Setting a too long or too short threshold for the time length of two or more phonemes obtained by adding the time lengths of the phonemes,
By subtracting likelihood points when the threshold is outside the range, the certainty of segmentation is reflected in the similarity degree, thereby achieving the effect of obtaining a high word recognition rate. (Description of Embodiment) An embodiment of the present invention will be described below with reference to the drawings. The basic configuration for carrying out the method of this embodiment is shown in FIG. 1, as in the conventional example. In FIG. 1, the word dictionary is the same as in the conventional example. The operation of this embodiment will be explained. First, the parameter extraction unit 1 analyzes input audio every 10 ms frame, extracts parameters, and creates a parameter time series. The process up to this point is the same as the conventional example. Next, this is compared with the word dictionary in the word dictionary unit 3, and for each dictionary item, phoneme segmentation is performed in the word recognition unit 2 according to the dictionary phoneme series that constitutes that dictionary item. The likelihood is calculated, which is a measure of the probability that the phoneme was uttered in the voiced segment. In this embodiment, the likelihood l _i of the i-th phoneme in the dictionary phoneme sequence is calculated by the following equation. l _i =l _i0 -l _ip1 -l _ip2 ... Here, l _i0 and l _ip1 are the same as in the formula of the conventional example. l _ip2 is the sum of the time length τ _i of the segmented interval for this i-th phoneme and the time length τ _i-1 of the (i-1)th phoneme is within a predetermined threshold. In Case of
If l _ip2 = 0, the value will be between 1 and 5000, and points will be deducted from the likelihood. For example (j−
1) If the th phoneme is G and the ith phoneme is A,
Table 1 shows the relationship between the threshold and l _ip2 . In Table 1, the threshold is set to two levels, one for too long and one for too short, and the value of l _ip2 is changed accordingly.

[Table] In other words, (τ _i + τ _i-1 ) that is actually impossible
, the likelihood is felt to be large, and such segmentation prevents a large degree of similarity from being obtained. After performing the likelihood calculation in this manner, the similarity is determined in the same way as in the conventional method. The effects of this embodiment will be described with examples. Shown in Figure 2. In the same input as the conventional example,
When the dictionary item was TOJONAKA, segmentation and likelihood calculation were performed normally as in the conventional example. Furthermore, l _ip2 = 0 for all phonemes in TOJONAKA. Dictionary entry is JAMAGATA
In the case of , the segmentation from the beginning of the word to G (i=5), the likelihood calculation, and the segmentation of the next A (i=6) were performed in the same way as before, but as mentioned above, τ ₅ =2 frame, τ ₆ = 1 frame, so l _6p2 = 4000 as shown in Table 1, and the likelihood of A is 4000 points lower than before. As a result, the similarity of JAMAGATA is also small,
The input was correctly recognized as TOJONAKA. In this example, when a segmentation that is clearly incorrect or rarely exists, points are deducted from the likelihood of the phoneme to reduce the similarity, and such segmentation is performed. This method has the advantage that it is possible to reduce the possibility that a dictionary item that has been edited is used as a recognized word, and to reduce misrecognition of words. In this embodiment, an example of an input word in which only one word is uttered is shown, but the same effect can be obtained for continuous words or words in a sentence. (Effects of the Invention) The present invention has the above-described configuration, and provides the following effects. When calculating the likelihood of a phoneme, the two phonemes obtained by adding the time length of one or more preceding phonemes to the time length of the segmented interval corresponding to that phoneme, or two phonemes. We set a threshold that is too long or too short for the above phoneme time length, and when it falls outside the range of the threshold, we feel the value of the likelihood, and such segmentation is impossible or rare. It is possible to reduce misrecognition of words so that a greater degree of similarity is not obtained, and improve the word recognition rate.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the basic configuration of an apparatus for implementing a speech recognition method in a conventional example and an embodiment of the present invention. FIG. 2 is a diagram showing segmentation in a conventional example and an embodiment of the present invention. 1...Parameter extraction unit, 2...Word recognition unit,
3...Word dictionary section.

Claims (1)

[Claims] 1. The input speech is checked against each dictionary entry in a word dictionary in which the word to be recognized is expressed in phonemes, and the input speech is segmented for each phoneme according to the dictionary phoneme series that constitutes each dictionary entry. The likelihood, which is a measure of the probability that the segment of voiced speech is the one that uttered the phoneme, is calculated, and this likelihood value is used to calculate the similarity between each dictionary item and the input speech and input it. In recognizing words,
When calculating the likelihood of a phoneme, calculate the length for the time length of two or more phonemes obtained by adding the time length of the phoneme interval to the time length of one or more phonemes before the phoneme. A speech recognition method characterized by setting a threshold that is too high or too short, and reducing the likelihood of the phoneme when the range of the threshold is exceeded.