JPS61177499A - 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 1. Gotachimi The present invention provides a speech section detection method for detecting a speech section used in a speech recognition device, more specifically, a method for determining an optimal threshold related to a threshold for cutting out a speech section. Regarding.

In general, in a speech recognition device, when the target is speech with a good signal-to-noise ratio, it is relatively easy to extract a section where speech exists. but,
The environment in which a speech recognition device is actually used includes various types of noise, and speech is input in a form superimposed on the noise. Since the noise at this time changes from moment to moment, it is difficult to detect a stable voice section using a method of setting a fixed threshold value and cutting out voice sections, which may cause misrecognition.

When clipping is performed using such a fixed threshold, low-power parts such as the beginning and end of speech and voiceless consonants are cut. Furthermore, under high noise conditions, noise will be added and cut out before and after what should normally be a voice section.

FIG. 6 is a diagram showing an example of a conventional voice section detection method.
a) The figure shows the audio power level, and the detection of the starting point St of the audio is performed at 1 d j - d j - 11>Ts, and the previous and current samples for each channel of the bandpass filter used for audio analysis are detected. If the sum of the absolute values of the power levels of By comparing the power level at the added threshold value, detect the voice section pulse signal as shown in (C), and then detect the -s pulse signal as shown in (b) by determining the length of the silent section (200 to 350 m5).
Generate an it pulse. However, in this method, as shown in Fig. 7, when the surrounding noise increases, the amplitude and extension level of the noise increases, leading to false detection in the previous start detection, and the threshold value TMI of the voice section also increases by a fixed amount. If q is small, there is a possibility of erroneously detecting noise.

The present invention has been made in order to solve the above-mentioned problems, and in particular, it is aimed at providing a speech segment detection method that can stably detect speech segments even under noisy conditions in a speech recognition device. -It is something that has been done.

In order to achieve the above object, the present invention provides a means for obtaining an audio power signal which is the sum of each channel, which has a bandpass filter for analyzing input audio, and a means for obtaining an audio power signal that is the sum of each channel, and a means for obtaining an audio power signal that is the sum of each channel, and a means for obtaining an audio power signal that is the sum of each channel. means for obtaining a differential power spectrum which is the sum of absolute values; means for detecting the start of a voice section by comparing the differential power spectrum with a certain threshold; and means for detecting the start of a voice section by comparing the previous voice power signal with a certain threshold; means for generating a binary voice interval, and a means for detecting a certain silent interval from the generated voice interval signal, and generating a one-word signal that uses the silent interval as a word delimiter when there is a silent interval of a predetermined length. In a speech section detection device having a means for generating a speech power signal and a means for storing a noise level of a differential power spectrum at the terminal edge of the single word pulse, the speech power signal is present at the time when the starting edge of the speech is detected. If it is above a threshold value, this is taken as the starting point, and after that, it is compared with the speech power signal at that threshold value to cut out the speech section.At the end edge of a single word pulse, a differential power spectrum for starting point detection and a difference power spectrum for subsequent section detection are used. It is characterized by storing the noise level of the voice power signal for the purpose of the present invention, and adding a certain coefficient times the value and a fixed value to each of the noise levels to obtain the threshold for detecting the start of the next word and the threshold for detecting the section that follows. This is what I did. Hereinafter, the present invention will be explained based on examples.

Generally, as the noise level increases, the noise power NP and its differential power spectrum DP increase as shown in FIG. 8, and it is desirable to change the thresholds of the starting point and voice section along this trajectory.

Therefore, in the present invention, the threshold values for the start point of the voice and the voice section are controlled as follows.

FIG. 1 is a diagram showing an example of an electric circuit for implementing the present invention, and FIG. 2 is a time chart for explaining its operation. In the figure, 1 is a microphone, 2 is a preprocessing section,
3 is a group of band pass filters (B, P, F), 4 is a recognition processing section, 5 is an audio power signal generation section, 6 is a differential power spectrum signal generation section, 7.8 is a comparison section, 9 and 10 are sample hold part, 11 is the difference power spectrum d to 1
12 is a calculation unit that doubles the audio signal power p and adds a fixed value q; 14 is a pulse control unit; 15 is a result output unit; The one-word pulse shown in Fig. 2 (b) is at point b, and the second pulse is at point 0.
The interval pulse shown in Figure (c) appears, the end edge signal appears at point d, and the start pulse signal appears at point e.

First, for each frame sample, calculate the difference power spectrum from the previous sample, and add a fixed value e to the difference power spectrum d at the end of the utterance, similar to the threshold for section detection, and then Use this as the threshold for detecting the beginning of the sample.

Next, the point e stored by the terminal edge is multiplied by (1
<K<2), so that no matter which point is memorized, it will be above that point. In this way, such a drawback is eliminated, and the value changes in accordance with changes in the noise level, and the fixed portion is fixed at a value that guarantees only the value when the noise level is small. In other words, 5 = K 1 d + e for start edge detection, and T! for voice section! 2p+q is set as the threshold for detecting the next utterance section, and control is performed to follow the surrounding noise level. Next, this start edge detection uses a differential power spectrum, but this method is particularly effective when the stationary noise level is high; on the other hand, when the noise level is low, it is better to perform start edge detection using the audio power threshold. can cut out the voice section well. In other words, when the noise level is high, it is difficult to detect changes in low-level speech that are hidden in the noise, and it is better to use the speech power signal to catch the beginning of a gradual change in a consonant. becomes. Therefore, if the difference power spectrum is used to detect the start end when the noise level is above a certain threshold value, and the audio power signal is used to detect the start end when the noise level is low, good section segmentation can be realized.

Next, in the above-described method, since a single point of noise level is stored, the threshold value may deviate greatly depending on the value. Therefore, as shown in FIG. 2(a), it is preferable to use the average value of the past n frames from the terminal edge.

FIG. 3 is an electrical circuit diagram showing an example of using the average value of the past n frames from the terminal edge.
1.22 is a peak hold circuit, 25.26 is an average value circuit, 27.28 is a comparison circuit, and other parts that function in the same way as in Figure 1 are given the same reference numbers as in Figure 1. It is. Therefore, in this embodiment, a peak hold circuit is also used, the maximum peak value and the average value within the n frames are used, and the maximum peak value is If it is smaller, use it as the threshold for the next word. In this way, it is possible to detect a section that passes through the noise bell without raising the threshold value unnecessarily. Of course, this applies to both the differential power spectrum for start edge detection and the audio power signal for section detection. In this case, as shown in FIG. 4, there are cases where the peak value becomes larger in the past n frames due to pulse noise or the like. In such a situation, the average value and subsequent calculation values will also rise relative to the actual noise level due to the peak value.

Therefore, in order to reduce this, when the average value and the peak value exceed a certain ratio, the above calculation formula is switched and the average value is decreased according to that ratio. .

FIG. 5 is a diagram showing an example of an electric circuit for this purpose, and the opening corresponds to part 30 in FIG. The arithmetic circuit 11 shown in Fig. 3 consists of two arithmetic circuits 11a and llb, and as mentioned above, when the noise average value and the peak value exceed a certain ratio, the arithmetic circuits 11a and llb are switched to perform the previous calculation. The calculation formula is changed to reduce the increase from the average value, thereby preventing unnecessary increases in the threshold value.

As is clear from the above description, according to the present invention, it is possible to follow the level of ambient noise and detect the optimal speech segment extraction threshold for the noise level under the noise.

[Brief explanation of drawings]

FIG. 1 is an electric circuit diagram for explaining an embodiment of the voice section detection method according to the present invention, FIG. 2 is a time chart for explaining the operation of the circuit in FIG. 1, and FIG. FIG. 4 is an electric circuit diagram showing another embodiment of the present invention; FIG. 4 is a diagram showing an example of pulse noise; FIG. 5 is an electric circuit diagram for explaining another embodiment of the present invention; 8 are diagrams for explaining an example of a conventional voice section detection method. DESCRIPTION OF SYMBOLS 1... Microphone, 2... Preprocessing section, 3... Band pass filter (B, P, F) group, 4... Recognition processing section. 5... Audio power signal generation section, 6... Differential power spectrum signal generation section, 7.8... Comparison section, 9.10.
...Sample hold section, 11...Arithmetic section that multiplies the differential power spectrum d by 1 and adds a fixed value e. 12... Arithmetic unit that doubles the audio signal power p and adds a fixed value q, 14... Pulse control unit, 15... Result output unit, 21.22... Peak hold circuit. 23.24...Sample hold circuit, 25.26.
...Average value circuit, 27.28...Comparison circuit, 31...
- Divide circuit, 32... Comparison circuit.

Claims (5)

[Claims] (1) A means for obtaining an audio power signal that is the sum of each channel having a band-pass filter that analyzes input audio, and a differential power that is the sum of the absolute values of the spectrum of each channel with one sample before. A means for obtaining a spectrum, a means for comparing the differential power spectrum with a certain threshold value to detect the start end of a voice section, and a means for comparing the previous voice power signal with a certain threshold value to generate a binary voice section after the start point. means for detecting a certain silent period from the generated speech period signal and generating a one-word signal that uses the silent period as a word break when there is a certain silent period of a certain length; In a voice section detection device having means for storing a voice power signal and a noise level of a differential power spectrum at the terminal edge of a pulse, if the voice power signal is equal to or higher than a certain threshold at the time when the start edge of the voice is detected, this is detected. The difference power spectrum for detecting the start point and the noise level of the voice power signal for detecting the subsequent section at the end edge of a single word pulse in which the speech section is extracted by comparing it with the speech power signal using the threshold value. A voice section detection method characterized in that a certain coefficient times the value and a fixed value are added to each of the stored values to obtain a threshold for detecting the start of the next word and a threshold for detecting a subsequent section. (2) If the noise level of the stored audio power signal is below a certain threshold, the starting edge is not detected using the differential power spectrum signal, and the audio power signal is extracted from the starting edge to the ending edge using only the audio power signal. The voice section detection method described in (1). (3) Find the average value of the noise level of the past n frames and the peak value therein from the terminal edge of the one-word pulse,
The voice section detection method according to claim (1), characterized in that when the value obtained by adding the coefficient times the average value and the fixed portion exceeds the peak value, the peak value is used as the threshold value. (4) If the peak value and the average value are above a certain ratio, the coefficient of the average value is varied from the peak value to determine the threshold value below the average value. The voice section detection method described in paragraph (1). (5) A variable length speech section is added to the beginning and end of the speech section depending on the threshold level of the speech section. The speech interval detection method according to any one of the items.