JPH0458636B2 - - 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.

Configuration of conventional example and its problems A conventional speech recognition method will be explained with reference to drawings.
In the figure, the word dictionary represents all the words to be recognized using phonemes, such as Asahi, Yuki,
Yuki is written as /ASAHI/, /JUUKI/, /JUKI/, etc. Phonemes and standard patterns are created in advance for each phoneme through preliminary experiments.

Next, the operation of the above conventional example will be explained. Analyze the input audio every 10ms frame, extract the parameters, and create a parameter time series. The parameters are calculated in advance. Next, the degree of similarity is calculated for each dictionary item, and when calculating this degree of similarity, segmentation of phonemes is performed according to the dictionary phoneme sequence that makes up the dictionary item.
The likelihood, which is a measure of the probability that the segmented speech interval is the one that uttered the phoneme, is calculated, and the similarity is calculated as the average value of the likelihood of each phoneme in the dictionary entry. The dictionary item with the maximum degree is determined to be a recognized word.

Here, a long vowel, for example /JUUKI/'s /UU/
Since it is usually not possible to find the boundary between /U/ and /U/, segmentation is performed by grouping /UU/ together, and the likelihood is also calculated by grouping them together. Note that a long vowel like /UU/ above can be thought of as a single long vowel phoneme, but in this conventional example, it is treated as a series of two phonemes /U/. . Therefore, the likelihood value for two phonemes is also calculated.

In this conventional example, the likelihood li of the i-th phoneme in the dictionary phoneme sequence is expressed by the following equation.

li=l _iO −l _iP ... Here, l _iO is calculated as a measure representing how well the parameters in each frame in the segmented interval match the standard pattern of that phoneme. In addition, l _iP represents the reduction in likelihood when the segmented interval is too long or too short. For long vowels, the formula is slightly different because the likelihood is calculated for multiple phonemes as described above, but it is basically the same, and the likelihood is determined by the degree of matching with the standard pattern and the deduction of points depending on the length. . Here, the words /JUUKI/ and /JUKI/ differ only in whether /U/ is a long vowel or a short vowel. In order to identify such words, in this conventional example, when calculating the likelihood of a long vowel, /UU/ in this example, if the length of the interval of /UU/ is shorter than a predetermined threshold, the likelihood is On the other hand, when calculating the likelihood of a normal short vowel, /U/ in this example, if the length of the /U/ section is longer than the threshold, the likelihood is deducted.

However, the above conventional example had the following drawbacks. The input word is /JUUKI/,
If the dictionary phoneme sequence is also /JUUKI/, /J/
In the segmentation of , the part with large parameter variations is set as the boundary between /J/ and /U/, but the /J/ interval becomes extremely long and the /UU/ interval becomes correspondingly short. Often. This is because when a semi-vowel is followed by a long vowel, the parameter changes unique to semi-vowels last longer than when a short vowel follows. In other words, /JUU/'s /UU/ is clearly longer than /JU/'s /U/, but /JUU/'s /UU/
is not simply an extension of /U/, but /
This means that the properties of J/ have continued for a long time.
Therefore, if we segment /J/ as described above, the length of /JUU/ compared to /JU/ will be eaten up by the length of /J/, and /J/ will become longer. The /UU/ part, excluding the , is short for a long vowel. In such a case, for the input /JUUKI/, the dictionary entry /
The likelihood of /UU/ in JUUKI/ becomes low due to the point deduction being too short, while the likelihood of /UU/ in dictionary item /JUKI/ becomes low due to the point deduction.
Since points are not deducted for U/, the likelihood is high and the similarity is also greater for /JUKI/, which has the disadvantage that the word recognition result is mistaken for /JUKI/.

OBJECTS OF THE INVENTION The present invention is intended to eliminate the drawbacks of the above-mentioned conventional examples, and aims to improve the accuracy of similarity calculation, thereby improving the word recognition rate.

Structure of the Invention In order to achieve the above object, the present invention, when determining the appropriateness of the length of a long vowel or short vowel following a semi-vowel, uses the interval length of the long vowel or short vowel and the length of the preceding semi-vowel. By using the sum of interval lengths, even if the interval length of a long vowel is short, if the sum with the interval of the preceding semi-vowel is long, the likelihood will not be deducted due to the long vowel being too short; Even if the length is short, if the sum with the preceding half vowel interval is long, the likelihood is subtracted, which has the effect of improving the accuracy of likelihood and similarity calculations.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. The basic configuration of this embodiment is the same as that of the conventional example, and the word dictionary and standard pattern of phonemes are also the same as those of the conventional example.

The operation of this embodiment will be explained. The difference between this embodiment and the conventional example is the method of calculating the likelihood of a vowel following a semi-vowel, and since the other parts are the same, the method of calculating the likelihood will be described below. Note that when we say semi-vowels here, we include not only semi-vowels at the beginning of words or between vowels, but also semi-vowels in persistent consonants, such as /J/ such as Riya (/RJA/). Likelihood calculation is performed by segmenting the input speech according to the dictionary phoneme sequence, and calculating the likelihood using a formula or a formula similar to the formula for long vowels, as in the conventional example.
However, in this embodiment, the deduction of the likelihood, that is, the way l _iP in the formula is determined is different from the conventional example. In this example, in the likelihood calculation of a long vowel following a semi-vowel, the deduction for a long vowel that is too short is as follows:
A threshold is set for the sum of the lengths of the long vowel section and the preceding half vowel section, rather than the length of the long vowel only section, and this is done when the sum is shorter than the threshold. Furthermore, when calculating the likelihood of a short vowel following a semi-vowel, the deduction for an excessively long short vowel is not based on the length of the segment containing only that vowel, but on the sum of the length of the segment of the short vowel and the segment of the preceding semi-vowel. A threshold is set for this, and this is performed when the sum is longer than the threshold.

An example of the effect of this embodiment will be described. The case of the input word /JUUKI/ shown in the figure, which is the same as the conventional example, will be described. In this input example, the dictionary entry is also /
When JUUKI/, the segmentation results are similar to the conventional example above, with the /J/ section being long and /UU/
The result was that the interval was short for a long vowel. However, in this example, /UU/ is
Since we look at the sum of the lengths including the section of J/, here the section of /J/ is long, so there was no deduction for the likelihood of being too short for /UU/. Also, the same input/
For JUUKI/, in the case of dictionary entry /JUKI/, the segmentation result is /J/ is /JUUKI/
Same section as /J/, /U/ is /JUUKI/
It became the same section as UU/. Here, the length of /U/ was the standard length for the short vowel /U/. However, in this example, the sum of the /U/ interval plus the /J/ interval is calculated, so the /J/ interval is very long, so the sum exceeds the threshold, and the /U/ interval is calculated. Likelihood was too long and points were deducted. As a result, for this /JUUKI/ input,
The degree of similarity in the dictionary item /JUUKI/ is greater than that in the conventional example because /UU/ is not deducted for being too short, while the degree of similarity in the dictionary item /JUUKI/ is greater than that in the conventional example. , /U/
The word recognition result was correct as /JUUKI/.

In this example, the likelihood deduction points for a long vowel that follows a semi-vowel that is too short, and the likelihood deduction points for a short vowel that follows a semi-vowel that is too long are calculated based on the length of the interval of those vowels and the interval of the preceding semi-vowel. By setting a threshold for the sum of the sums, there is an advantage that the likelihood of long vowels and short vowels that follow a semi-vowel can be accurately deducted, and the accuracy of likelihood calculation can be improved.

Effects of the Invention The present invention has the above-mentioned configuration, and limits the length of a long vowel that follows a semi-vowel and the length of a short vowel that follows a semi-vowel. This is done by setting a threshold for the sum of the lengths and subtracting the likelihood, which improves the accuracy of similarity calculation to increase the degree of discrimination between long and short vowels, resulting in a high word recognition rate. can be obtained.

[Brief explanation of the drawing]

The figure is a diagram showing a speech recognition method according to a conventional method and an embodiment of the present invention.

Claims (1)

[Claims] 1. The input speech is compared with dictionary entries in a word dictionary in which words to be recognized are expressed in phonemes, and the input speech is segmented for each phoneme according to the dictionary phoneme series that constitutes each dictionary entry. The system calculates the likelihood, which is a measure of the probability that the phoneme was generated in the segment of the recorded speech, and uses this likelihood value to find the degree of similarity between each dictionary entry and the input speech. When recognizing an input word, limit whether a long vowel following a semi-vowel is too short or a short vowel is too long based on the sum of the lengths of that long vowel or short vowel and the preceding semi-vowel. A speech recognition method characterized by: