JPS62244100A - 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] First summer The present invention relates to a speech segment detection method for a speech recognition device.

In the detection of voice segments by the Inmai Engineering speech recognition device, detection of the beginning of a word has an important meaning, and it is especially important to accurately detect when the beginning of a word starts with a consonant. This starting point detection rarely causes detection errors in a good environment, but in an actual field, the noise level is high and it is easy to detect the consonant part at the beginning of a word. In reality, the start point is often detected based on the magnitude of the differential power spectrum, but in many cases it is not possible to successfully extract the consonant part that gradually rises using this method alone. In addition, in the past, methods were used in which a voice section is defined as a voice section equal to or above a certain threshold of voice power, and a method in which a differential power spectrum is used only to detect the beginning of the voice.These methods work well when the surrounding environment is very quiet, but When stationary noise or the like is present, the threshold value generally increases in conjunction with the noise, making it easier for consonant parts to be gradually dropped.

FIG. 5 is an electrical block diagram for explaining an example of starting edge detection using a differential power spectrum.
In the figure, 1 is a microphone, 2 is a preprocessing unit, 3 is a bandpass filter, 4 is a voice power generation unit, 5 is a section comparison unit, 6 is a differential power generation unit, 7 is a start end comparison unit, and 8 is a voice section generation unit. In the section, the audio input from the microphone 1 is frequency-analyzed by sampling at a certain period by the band-pass filter group 3, and then the channel power of each channel is output. In the differential power generating section 6, the differential values from each channel power one sample before are totaled to form a differential power spectrum. If this differential power spectrum exceeds a certain threshold, it is determined as a voice start point, and after the start point is determined, the voice power is monitored, and a voice section that is equal to or higher than a certain threshold is determined as a voice section.

FIGS. 6 and 7 are waveform diagrams for explaining examples of sound interval detection determined as described above.

In both figures, (a) shows the audio power, and (b) shows the differential power. Thus, as shown in Figure 6.

When the sound starts from a vowel, it is extracted with a relatively small error, but as shown in Figure 7, when the voice power rises gradually, when the sound cannot be extracted at point (A), (
It becomes impossible to detect the beginning up to point (b), and the consonant block at the beginning of the word is simply dropped. In particular, /s/l
/p/l There are many words like this that start with /l/, /M/, '/N/, etc.

However, if the threshold value of the differential power spectrum is reduced in order to extract these starting edges, or if the starting edges are extracted using audio power, the system becomes vulnerable to stationary noise and environmental noise, making the system difficult to use in actual use.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, this was done for the purpose of more accurately detecting speech sections in a speech recognition device.

Configuration In order to achieve the above object, the present invention provides means for detecting the audio power of audio input from a microphone, n (n =
In a speech interval detection method having a bandpass filter for channels (integer), means for analyzing five frequencies to obtain the power of each channel, and means for detecting a frequency difference power spectrum of a sampling period, in detecting the start of speech. When the audio power in the high-frequency range is greater than the audio power in the low-frequency range, it is determined as the start of a voice section, or in detecting the start of the voice, the audio power in the low-frequency range is higher than the voice in the high-frequency range. Processing to set the start of a voice section when the power is greater than the voice power, or to set the start of a voice section when the voice power in the high frequency range is greater than the voice power in the low frequency range. Processing that sets the time when the audio power of the part is larger than the audio power of the high frequency part as the start of the audio section, and
This process is characterized in that the start of the voice is defined as the start of the voice when the differential power spectrum exceeds a certain threshold value, and the start of the voice is determined when any of the conditions are met. Hereinafter, the present invention will be explained based on examples.

The present invention aims to overcome the above-mentioned drawbacks and stably detect consonant blocks.

When separating /s/, /p/, /l/, /M/, /N/, etc. from noise, it is desirable to focus on the frequency components, and generally, /s / e / p / */
1/ has high-frequency power, and /M/, /N/, etc. have power concentrated in the low-frequency range. Therefore, the output of the bandpass filter is divided into three powers: high-frequency, middle-frequency, and low-frequency, and the respective powers are compared only when detecting the starting edge.

FIG. 1 is an electrical block diagram for explaining an embodiment of the present invention, in which 11 is a microphone, 12 is a preprocessing section, 13 is a group of bandpass filters, and 14 is an audio power generation section. , 15 is a high frequency power generation section, 16 is a middle frequency power generation section, 17 is a low frequency power generation section, 18 is a differential power generation section, 19 is a comparison section, 20 is a power comparison section, 21 is an interval generation section, First, in order to detect /s/, /p/, /l/, etc., when the high frequency power is greater than a certain threshold (Thigh) and is much larger than the low frequency, it is /S/.

It is determined that a consonant such as /p/ or /l/ has started.

In most cases, this condition means that there is almost no power in the low range, and energy is concentrated only in the high range, making it easy to distinguish noise etc., and conversely, /M/, /N /
etc., the starting point is determined when the power is greater than the threshold (T10) and is much larger than the high frequency power, but in this case, the energy is concentrated in the low frequency range and does not exist in the high frequency range. . Figure 2 shows this situation.If the above two conditions are met, it is immediately set as the starting point, and if the conditions are not met, the differential power spectrum is checked.If the OR condition is set, all words Good detection is possible.

FIG. 3 is an electrical block diagram for explaining another embodiment of the present invention. In the figure, numeral 21 is a flip-flop circuit, and the other functions are similar to those of the embodiment shown in FIG. 1. The parts are given the same reference numerals as in FIG.

In the embodiment shown in FIG. 1, when generating sections under the above conditions, in order to detect /M/, IN7, etc.
It becomes easier to detect a buzz sound immediately before a voice is uttered, and on the other hand, there is a possibility that unnecessary information is added to a voice section. As shown in FIG. 4, this buzz sound is characterized by a stronger concentration in the low range and a lower power level than /M/, 7N/, etc.

In order to eliminate this buzz sound, in this embodiment, the low range power is compared with the power of other mid ranges and high ranges, and if the low range power is much larger than the other mid range and high range powers, and the low frequency power is smaller than a certain dark value (Tbazz), and even if the above-mentioned starting point condition is met, the starting point is not set as long as this condition is satisfied. .

Effects As is clear from the above explanation, according to the present invention, it is possible to more stably detect the start of a voice section.

[Brief explanation of drawings]

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, FIG. 2 is a diagram for explaining the operation of the embodiment shown in FIG. 1, and FIG. 3 is an electrical block diagram for explaining one embodiment of the present invention. 4 is a diagram for explaining the operation of the embodiment shown in FIG. 3, and FIG. 5 is an electrical block diagram for explaining another embodiment of the invention. An electrical block diagram for explaining an example of detection, FIGS. 6 and 7, are diagrams for explaining the operation of the electrical block diagram shown in FIG. 5, respectively. 11... Microphone, 12... Preprocessing section, 13
...Band pass filter group, 14...Audio power generation section, 15...High frequency power generation section, 16...Medium frequency power generation section, 17...Low frequency power generation section, 18...・
Difference power generation section, 19... Comparison section, 20... Power comparison section, 21... Section generation section. Figure l Figure 2 @3I! ! 11 Fig. 4 □Honosuke

Claims (4)

[Claims] (1) means for detecting the audio power of the audio input from the microphone; means having an n (n = integer) channel bandpass filter and analyzing the frequency to obtain the power of each channel; and a sampling period. In a voice section detection method having means for detecting a frequency difference power spectrum of A speech interval detection method characterized by the following. (2) means for detecting the audio power of the audio input from the microphone; means having an n (n = integer) channel bandpass filter and analyzing the frequency to obtain the power of each channel; and a sampling period. In a voice section detection method having means for detecting a frequency difference power spectrum of A speech interval detection method characterized by the following. (3) In detecting the start of a voice, when the voice power in the low frequency range is greater than the voice power in other parts, and when the voice power in the low frequency range is less than a certain threshold, the claim Claim (2) characterized in that the start edge detected in the range (2) is canceled.
The speech interval detection method described in Section. (4) means for detecting the audio power of the audio input from the microphone; means having an n (n = integer) channel bandpass filter and analyzing the frequency to obtain the power of each channel; and a sampling period. In a voice section detection method that has means for detecting a frequency difference power spectrum of Either one of the following conditions is met: a process in which the start of a voice section is defined as the time when the voice power is greater than the voice power in the high frequency range, or a process in which the time when the differential power spectrum is greater than a certain threshold is determined as the start of the voice section. A voice section detection method that is characterized in that it is determined as the beginning of a voice when it occurs.