JPS63306497A - Voice section detecting system - 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] Technical field The present invention relates to a speech segment detection method for a speech recognition device.

雛]θE4 In conventional voice section detection, the end of the voice section at the end of a word is defined as the point where the voice power becomes smaller than a certain value.

However, with this method, there is little error for vowels where the power at the end of the first word suddenly decreases, but for the "n" sound where the power at the end of the word slowly decreases, it is often considered to be the end in the middle of the voice. Because the "n" sound has low power, in some cases, the "n" sound was missing.

1-1 The present invention was made in view of the above-mentioned circumstances.
In particular, this method was developed for the purpose of correctly detecting a segment of a word ending in "n" in a speech segment detection method.

Configuration In order to achieve the above object, the present invention provides means for extracting features of input speech at a certain sampling period, means for determining a speech section using power, and means for extracting features of input speech at a certain sampling period, means for determining a speech section using power, and a means for extracting features of input speech at a certain sampling period. and a means for detecting the vocal part of "n" by comparing the power, and a means for detecting the vocal part of "n" to determine whether the end of one word ends with "n". If the sound ends with "n", the voice after the end of the section is searched and the state of "n" continues and the power is 1/2 of the vowel power in the word.
This feature is characterized in that if it is larger than a predetermined value of ~1/16, it is considered a voice section. An explanation will be given below based on one embodiment of the present invention.

FIG. 1 is a time chart for explaining one embodiment of the present invention. For example, when the input voice is ``han'', the voice power is as shown in FIG. The part shows a stable level, and the power of "n" is small, and the power tends to gradually decrease. Therefore, if we define the part with power above a certain value as a vocal section, it will become as shown in Figure 1 (b), where most of the part ``n'' will disappear; If the threshold value is set small so that it is considered as , section extraction becomes impossible if noise is added. Therefore, in the present invention, only words whose endings end in "n" are corrected to correct the speech interval, thereby correctly cutting out the interval. ``H'' is detected by comparing the high frequency component power H (approximately 800 Hz) and the low frequency component power L (approximately 800 1z or less).

Figure 1 (c) shows the respective powers for "han", and as shown in Figure 1 (d), the section where the low frequency component power L is greater than the high frequency component power H is for "n". It can be considered as an interval. Therefore, search q from point C, which is about 100+s back from the end of the voice section in FIG. In the figure, since the end of the word ends with "n", section correction is performed.If the ending of the word does not end with "n", this condition does not apply, so the phonetic section shown in Figure 1 (b)
1) is the voice section. Correction is performed by adding an interval where the following two conditions are both satisfied.

(1) As shown in FIG. 1(d), the "n" section continues.

(2), as shown in Figure 1 (e), the vowel power is detected (ends at point (E)), and then the power (■) is about 1/8 of the vowel level compared to the voice power 5. A large thing (value is determined from 1/2 to 1/16).

The reason for (1) above is to omit sections where the feature amount is inaccurate due to noise etc., and the reason for (2) is that the feature amount becomes inaccurate if the level is too small. (d) in Figure 1
The section that satisfies the AND condition of (e) is the 1st rJ! 1(
By adding to b).

A corrected speech section (2) is created (Fig. 2(f)),
The "n" section is correctly detected from the voice section (1) in the normal section of FIG. 2(b).

FIG. 2 is a block diagram for explaining an embodiment for realizing the voice section detection method of the present invention.

In the figure, 1 is a microphone, 2 is an amplifier, 3 is a feature extraction unit, 4 is a voice power detection unit, 5 is a high frequency component power detection unit, 6 is a low frequency power detection unit, 7 is a vowel power detection unit, and 8 is a vowel power detection unit. Voice section threshold section, 9 to 11 are comparators, 12
is the word ending [n] detection part.

Reference numeral 13 denotes a section correction section, in which audio input from the microphone 1 is amplified by an amplifier 2, and a feature extraction section 3. Audio power detection section 4
, high frequency component power detection section 5. The 0 normal speech section (speech section (1)) input to the low frequency component power detection section 6 and the vowel power detection section 7 is obtained by comparing the speech power and the speech section threshold value by the comparator 9. Also, the “n” section (
FIG. 1(d)) is obtained by comparing the high frequency and low frequency component powers using the comparator 10. The ending "n" detection unit 12 detects whether the ending of the word ends with "n" from this voice section signal and the "n J section signal.

If so, the speech section is corrected using the AND condition of the [n] section and the section having a power of 1/8 or more of the vowel output from the comparator 11, and the speech section shown in FIG. 1(f) is created. If the ending of the zero word to obtain (2) is "n", no correction of the speech interval is performed.

As is clear from the above explanation, according to the present invention, it is possible to detect the correct speech section of a word whose first word ends in "n".

[Brief explanation of the drawing]

FIG. 1 is a time chart for explaining an embodiment of the present invention, and FIG. 2 is a configuration diagram for explaining an embodiment for realizing the present invention. 1...Microphone, 2...Amplifier, 3...Feature quantity extraction unit, 410. Audio power detection section, 5... High frequency component power detection section, 6... Low frequency power detection section, 7
...Vowel power detection section, 8...Speech interval threshold section, 9
~11... Comparator, 12... Word ending "n" detection unit, 13... Section correction unit. Figure 1 Sound power

Claims (1)

[Claims] A method for extracting features of input speech at a certain sampling period, a method for determining speech sections using power, and a method for detecting the speech part of "n" by comparing the power of high frequency components and the power of low frequency components. In a voice section detection method having a means for detecting a sound, it is determined whether a word ends with "n" using the above-mentioned means for detecting the sound part of "n", and if it ends with "n". Search for the audio after the end of the section,
A voice section detection method characterized in that if the state of "n" continues and the power is greater than a predetermined value of 1/2 to 1/16 of the intra-word vowel power, it is determined as a voice section.