JPH045400B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) 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.

(Structure of conventional example and its problems) Fig. 1 is a block diagram showing an outline of the functions of a device that implements one conventional word recognition method (first conventional example), and Fig. FIG. 7 is a diagram showing a change in the ratio of the mid-range band power and the low-band power in the W interval.

In FIG. 1, 1 is a parameter extraction section, 2 is a phoneme segmentation section, 3 is a phoneme recognition section, and 4 is a phoneme recognition section.
5 is a word dictionary section, 5 is a confusion matrix section, and 6 is a word recognition section. The word dictionary section 4 stores a word dictionary in which all words to be recognized are expressed in phonemes, and the word dictionary includes, for example, the words "Satsuporo", "Asahikawa", "Wakayama", "Okayama", etc. , "SAQPORO", "ASAHIKAWA",
It is written as "WAKAJAMA", "OKAJMA", etc. The confusion matrix unit 5 shows the probability that various phonemes used in dictionary notation will be recognized in actual phoneme recognition, for example, the probability that A will be recognized as A is 85%, and the probability that A will be recognized as O is 85%. The probability is 7
%, the probability that A will be dropped from the recognized phoneme sequence due to a segmentation error is 5%, and so on.

Next, the operation of the above conventional example will be explained. The parameter extraction unit 1 analyzes the input audio every 10ms frame and extracts the parameters.The phoneme segmentation unit 2 performs phoneme segmentation (input audio The phoneme recognition unit 3 performs phoneme recognition for each segmented interval. Word recognition unit 6
Now, the degree of similarity between the recognized phoneme sequence obtained by the phoneme recognition unit 3 and each dictionary item stored in the word dictionary unit 6 is calculated using the confusion matrix in the confusion matrix unit 5, The word (dictionary entry) with the highest degree of similarity was selected as the recognized word. The segmentation method for the Wa vowel (hereinafter referred to as W) in this conventional example is based on the band power P M in the middle range (600 to 1500 Hz) of the input voice and the band power P _M in the low range (250 Hz).
~600Hz), the ratio R _ML R _ML △=P _M /P _L shown in the formula below is divided into segments with _the valley-shaped section as the section of W as shown in Figure 2. It was something to look forward to.
This utilizes the property that the _RML changes in a valley shape in response to the change in the first formant frequency in the W interval, but it has the following drawbacks.
In other words, since R _ML is a quantity that fluctuates depending on the change in formant frequency even outside the W interval, the valley of R _ML occurs even in intervals other than the W interval, and the depth of the valley of R _ML must be adjusted in order not to mistake that interval for the W interval. In this case, a certain threshold was set, and when the depth of a valley exceeded that threshold, that valley section was segmented as the W section, and other shallow valleys were ignored. As a result, only about 70% of the W sections are correctly segmented, and there are many erroneous word recognitions. For example, the first
As shown in the example in the figure, the input word is WAKAJAMA
, this W is often not segmented for the reasons mentioned above, and the obtained recognized phoneme sequence becomes AKAJAMA, and as a result of calculating the similarity with each dictionary entry in the word dictionary, the word recognition result is
I made a mistake with OKAJAMA. This conventional example has a drawback in that there are many erroneous word recognitions as shown above.

Next, a second conventional example will be described with reference to FIG. In FIG. 3, the parameter extraction section 1 and word dictionary section 4 are similar to those in the first conventional example shown in FIG. In the second conventional example, the parameter extraction unit 1 analyzes the input every 10 ms frame and extracts the parameters, and the word writing unit 7 directly uses the parameter time series obtained by the parameter extraction unit 1.
The word dictionary of the word dictionary section 4 is compared, and for each dictionary item, assuming that the input speech is the one that uttered that dictionary item, segmentation is performed one phoneme at a time according to the dictionary phoneme sequence of that dictionary item. , calculates the likelihood, which is a measure of the probability that the segmented interval is the utterance of that phoneme in the dictionary phoneme sequence, and calculates the likelihood, which is a measure of the probability that the segmented interval is the utterance of that phoneme in the dictionary phoneme sequence, and uses the average value of the likelihood as the difference between the dictionary entry and the input speech. Words were recognized by determining their similarity. Here, in the second conventional example, when performing segmentation of the W interval corresponding to W on the dictionary phoneme sequence, this distance is the closest among the distances to each vowel standard pattern found for each frame. The W interval was segmented by utilizing the property that the vowel type changes from U to A or O to A in ``wa''. However, in the second conventional example, unless the utterance is very clear, the closest vowel in each frame in the W section is often the final A or O, and the change cannot be detected. In this case, segmentation of the W interval is not possible, that is, it tends to be determined that the W interval is not included, leading to word recognition errors.

(Object of the Invention) The present invention eliminates the drawbacks of the conventional example, and aims to improve the word recognition rate by correctly segmenting the W section.

(Structure of the Invention) In order to achieve the above object, 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. At the same time, when segmenting the Wa-line semi-vowel, the ratio of the mid-range band power P _M and the low-range band power _PL of the input voice is determined.
The time change of R _ML is also compared, and when the ratio R _ML is the minimum, segmentation is performed as a Wa line semi-vowel. After segmentation, the likelihood of each phoneme is calculated, and this likelihood is calculated. This is a speech recognition method that recognizes input words by determining the degree of similarity between each dictionary item and input speech using the value of . Due to this feature, the present invention can perform segmentation reliably and has 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 of an apparatus for carrying out the speech recognition method of this embodiment is shown in the block diagram of FIG. 3, as in the second conventional example. In FIG. 3, the parameter extraction section 1 and the word dictionary section 4 are the same as those in the first conventional example and the second conventional example.

The operation of this embodiment will be explained. First, the parameter extraction unit 1 analyzes the input voice every 10 ms frame and extracts parameters, and the word recognition unit 7 directly compares the parameter time series obtained by the parameter extraction unit 1 with the word dictionary of the word dictionary unit 4. For each dictionary item, assuming that the input speech is the one that uttered that dictionary item, segmentation is performed one phoneme at a time according to the dictionary phoneme sequence of that dictionary item, and the segmented interval is calculates the likelihood, which is a measure of the probability that the phoneme in the dictionary phoneme sequence was generated,
A word is recognized by determining the similarity between the dictionary entry and the input speech as the average value of the likelihood. At this time, in this embodiment, when performing segmentation of the W section using the W-line semi-vowel W on the dictionary phoneme sequence, the mid-range band power and the low-range band power of the input voice shown in equation (1) above are calculated. Segmentation is performed by setting the section where the power ratio _RML has a valley shape as shown in FIG. 2 as the section W.

In this embodiment, even if a valley in R _ML occurs in a place other than the W interval, there is no problem at all as long as the W segmentation of the dictionary entry different from the input word is not performed in that part. Even if such incorrect segmentation is performed, it can be distinguished from the correct W interval segmentation by performing likelihood calculations that also use other parameters so that the likelihood is low. . Therefore, even if the valley of R _ML is used for segmentation of W interval,
Unlike the first conventional example, there is no need to set a threshold for the depth of the valley, and even if the _RML valley is shallow, the W section can be correctly segmented, and the word recognition rate is improved. In addition, since R _ML rarely forms a valley in the W section,
Compared to the second conventional example, the rate at which the W interval can be correctly segmented is much higher, and the word recognition rate is also improved. For example, as shown in Figure 3, the input word is
When WAKAJAMA, the segmentation of W is done reliably and the word recognition result is also correct.
It became WAKAJAMA.

(Effects of the Invention) The present invention has the above-described configuration, and provides the following effects. When input speech is segmented according to the dictionary phoneme sequence of each dictionary item, segmentation of W is defined as the ratio of the mid-range band power to the low-range band power of the input speech that becomes a valley in the W interval. By using W
This has the advantage that the sections are reliably segmented and the word recognition rate is improved.

[Brief explanation of drawings]

FIG. 1 is a block diagram schematically showing the functions of a device used to implement the first conventional speech recognition method, and FIG. 2 shows the ratio of mid-range band power to low-band power in the W section. FIG. 3 is a block diagram schematically showing the functions of the apparatus used to implement the speech recognition method in the second conventional example and the embodiment of the present invention. 1...Parameter extraction section, 4...Word dictionary section,
7...Word recognition section.

Claims (1)

[Claims] 1. The words to be recognized in the input speech are compared with each dictionary entry in a word dictionary expressed in phonemes, and the input speech is segmented for each phoneme according to the dictionary phoneme series that constitutes each dictionary entry. During segmentation, the time change of the ratio R _ML of the mid-range band power P _M and the low-range band power _PL of the input audio is also checked, and when the ratio R _ML is the minimum, the segmentation is performed. Segmentation is performed as a line semi-vowel, and then the likelihood, which is a measure of the probability that each segmented speech interval is the one that uttered that phoneme, is calculated, and this likelihood value is used to segment each segment. A speech recognition method characterized by recognizing words in input speech by determining the degree of similarity between dictionary entries and input speech.