JPS58129498A - Voice recognition - 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 speech recognition method, and an object of the present invention is to extract with high accuracy devoiced vowels that cannot be recognized by ordinary vowel recognition means, and to improve the speech recognition rate.

First, a conventional devoiced vowel extraction method will be described. The basic idea is to first perform rough phoneme recognition, and when an unvoiced interval is long and there is a significant voice power ditsuno between them, this D, f is regarded as the boundary between two syllables, and the unvoiced interval is recognized. It is recognized as a series of three phonemes: a voiceless consonant, a voiceless vowel, and a voiceless consonant. This will be explained in detail with reference to FIG. The input audio is divided into frames of every 10m5, analyzed at the acoustic level, and the -7 meters necessary for phoneme recognition are extracted. Next, segmentation (finding phoneme boundaries) and preliminary phoneme recognition are performed on a frame-by-frame basis. At this point, that's the only voiced vowel! ! However, in the case of devoiced vowels,
The phoneme including the consonants on both sides is recognized as one voiceless consonant. Therefore, if a voiceless consonant section is found as a result of preliminary phoneme recognition, it is necessary to check the presence of a voiceless vowel. The conventional check routine is represented by a diamond-shaped branch in FIG. First, check whether there are voiceless consonants. If it does not exist, no further checks are performed. Furthermore, if there is an unvoiced consonant section, the number of frames in this unvoiced consonant section is greater than a threshold Tll determined in advance based on experimental results, and a silent (represented by Q) frame exists in this unvoiced consonant section, Moreover, when the number of frames in this Q interval is greater than the predetermined threshold TI2, this Q interval is considered to be a boundary between two syllables, and the unvoiced consonant interval is divided into two halves, one in front and the other in half. Insert a vowel UI immediately after the voiceless consonant. This UI means the phoneme ``U'' or ``tech''. After performing the devoicing process as described above, a recognized phoneme sequence is created and matched with a word dictionary to perform word recognition.

In the conventional devoiced vowel extraction method described above, T12
If TI2 is set to a small value, the number of cases in which devoiced vowels are incorrectly extracted from a devoiced consonant interval that originally contains - devoiced consonants increases, and on the other hand, if TI2 is set too high, the extraction rate of devoiced vowels decreases. There was a drawback.

The present invention eliminates the drawbacks of the conventional example, and embodiments will be described below, focusing on the differences from the conventional example. FIG. 2 shows a recognition flow according to an embodiment of the present invention. It is only within the circle (201) indicated by the al line shown in FIG. In other words, even if the Q interval is short, the local maximum and minimum values of the difference value of Pwa (difference between frames) between the Q intervals are determined, and the difference is set as a predetermined threshold T23. If the Q interval is too large, it is assumed that there is a clear break in the voice, and the Q interval is treated in the same way as when the Q interval is long.

Explain the operation of the word "machine".

The ``ki'' in ``kikai'' is devoiced in standard Japanese.

Therefore, the preliminary recognition result of "Kikai" is not K I KA I, but ■ is omitted, and two K's are not lucky.

Figure 3 shows the preliminary recognition results for the first half of ``Kikai'' and the changes in ``Kikai'' with the same time axis.

The silent Q is recognized to overlap with the consonant K. This is because silent parts in words (more precisely, parts where the nokwa is below a predetermined threshold) are also caused by instability during consonant pronunciation. This is because it is not possible to determine whether it is a syllable boundary. The consonant K and the vowel A are recognized overlappingly by 1 frame.

This is because phoneme boundaries are currently recognized overlappingly, and there is no particular reason for this. Now, regarding the Q interval in Figure 3, if this Q interval is longer than a predetermined threshold or has a large change, it is considered a syllable boundary, and the first half K becomes K and UI, and the second half K becomes K connected to A, and the devoiced vowel has been extracted.

Note that when calculating power, the true power is 2 of the instantaneous value.
Although it is obtained by integrating the power over a certain period of time, almost the same result can be obtained even if the absolute value is used instead of the square.

The following table shows the results of recognition experiments conducted in the above embodiment and the conventional example.

In the above table, the input CvC shows only cases in which unvoiced consonants, unvoiced vowels, and unvoiced consonants are connected, and the number of correctly recognized CVCs and the number of unvoiced consonants recognized as one consonant are shown. Indicates the number of pieces.

The lower row shows the number of unvoiced consonants that were input, but the unvoiced vowel extraction method was applied incorrectly and the recognition result was CVC. In addition, various thresholds have been determined through preliminary experiments for both the conventional example and the example.
The optimum value was set. Looking at the results, the extraction rate of devoiced vowels was 715% in the conventional method, but improved to 987% in accordance with the present invention. On the other hand, in the past, incorrect application of the devoiced vowel extraction routine occurred 12% of the time when it was applied correctly, but with the present invention, this has been eliminated. Although the present invention does not directly aim at reducing the number of erroneous applications, it is possible to reduce the number of erroneous applications by tightening the threshold of feldspar in the Q section.

As described above, according to the present invention, the extraction rate of devoiced vowels has been greatly improved, while the erroneous extraction rate has been greatly reduced. Such an improvement in the recognition rate of devoiced vowels is directly linked to an improvement in the word recognition rate, and the effects of the present invention are significant.

[Brief explanation of the drawing]

Figure 1 is a flow diagram of a conventional speech recognition method, and Figure 2 is a flow diagram of a conventional speech recognition method.
FIG. 3, which is a flowchart of the speech recognition method according to an embodiment of the present invention, is a diagram showing an example of preliminary phoneme recognition results of input speech and changes in noise.

Claims (1)

[Claims] The unvoiced sound continues for longer than the normal length of one phoneme, and there is a voice power division within this unvoiced sound section, and this sound no. 9
When the rate of change of ・guwa in Wadizo with respect to time is larger than a predetermined value, the unvoiced sound section is ・, −
A speech recognition method characterized in that phoneme recognition is performed on the assumption that three phonetic elements consisting of a voiceless consonant, a voiceless vowel, and a voiceless consonant are in a continuous state instead of individual voiceless consonants.