JPS6250837B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an audio region detection device that detects the time domain of an audio signal from an input signal including the audio signal.

Such a voice region detection device is used in voice recognition equipment that can recognize voices, which has been actively developed recently.
This type of speech area detection device is shown in FIG. 1, which is used to create speech patterns. In the figure, 1 is a microphone that converts audio into an electrical audio signal, 2 is a microphone amplifier that amplifies the audio signal from the microphone 1, and 3 is a microphone that converts the audio signal obtained from the microphone amplifier 2 into an audio signal. 4 is a buffer memory that stores the frequency spectrum from the spectrum extractor 3 in time series; 5 is a first comparator that receives the input from the microphone amplifier 2; The peak value of the signal is compared with a first threshold (THR1), and when the peak value exceeds the first threshold (THR1), an audio input detection signal is output. Reference numeral 6 denotes a second comparison circuit, which compares the peak value of the input signal from the microphone amplifier 2 with a second threshold (THR2) that is smaller than the first threshold (THR1), and compares this peak value with the second threshold (THR2). A sound pressure drop detection signal is output when the sound pressure drops below the threshold (THR2). 7 is a counter circuit that counts the time during which the sound pressure drop detection signal is output from the second comparison circuit 6, and outputs an audio input end signal when this count reaches 150 msec. . Reference numeral 8 denotes a pattern editing circuit, which sets the time point before 50 msec from the time when the audio input detection signal from the first comparator 5 is received as the audio input start time, and also when the audio input end signal from the counter circuit 7 is received. A voice region is set with the time point at which voice input ends, and the frequency spectrum stored in the buffer memory 4 included in this voice region is read out. Reference numeral 9 denotes an audio pattern memory that stores the frequency spectrum read out by the pattern editing circuit 8.

The operation of the above-mentioned audio area detection device will be explained using the input signal waveform diagrams shown in FIGS. 2a and 2b. In the case of FIG.
If only the input signal is input, the time point (T ₀ ) 50 msec before the time point (T ₁ ) when the peak value of this input signal exceeds the first threshold value (THR1) of the first comparator 5 is the audio pattern editing circuit. 8, the voice input start point is set. On the other hand, during the period (T ₂ to T ₃ ) in which the peak value of the input signal is temporarily lower than the second threshold value (THR2) of the second comparator 6 within 150 msec, the counter circuit 7 An audio input end signal is not obtained, and after the time (T ₄ ), an audio input end signal is obtained from the counter circuit 7 at a time (T ₅ ) 150 msec after this time (T ₄ ). The time point (T ₅ ) is set by the voice pattern editing circuit 8 as the voice input end time point.

However, if the ambient noise N is input to the microphone 1 together with the audio signal S at a relatively high level as shown in FIG. Even if the peak value of the signal S decreases, the level of the ambient noise N is higher than the second threshold value (THR2) of the second comparison circuit 6, so it is no longer possible to set the end point of audio input, and the end point of the audio input cannot be set. It was preventing detection.

The present invention has been made in order to eliminate the above-mentioned disadvantages, and provides a voice region detection device capable of detecting a voice region even from an input signal with a high noise level.

FIG. 3 shows an example of the voice area detection device of the present invention for creating voice patterns. In the same figure,
1 to 4, and 8 and 9 indicate a microphone to a buffer memory, an audio pattern editing circuit, and an audio pattern memory, as in the conventional example of FIG. 1 comparator 5
It performs the same operation as . The difference between the sound region detection device according to the present invention and the conventional device is that the noise recognition means 10 is used instead of the second comparator 6 for detecting a drop in the peak value of the input signal to obtain a sound pressure drop detection signal. The point is that it has been established. The noise recognition means 10
receives the audio input detection signal from the comparator 5', and stores the frequency spectrum in the time domain even earlier than the point 50 msec before this signal is received, that is, the noise spectrum string based only on the noise signal, into the buffer memory 4. a noise pattern editing circuit 11 that derives a noise pattern obtained by reading out and averaging these noise spectrum sequences; a noise pattern memory 12 that immediately stores the noise pattern derived by the noise pattern editing circuit 11; a noise region detection circuit 13 that sends a noise region detection signal to the counter circuit 7' when the noise pattern in the noise pattern memory 12 matches each frequency spectrum obtained successively in time series from the spectrum extractor; And so on. The noise recognition means 1
Similar to the conventional counter circuit 7, the counter circuit 7', which counts the time during which the noise area detection signal from the noise area detection circuit 13 of 0 is continuously obtained, detects the sound when the counted value reaches 150 msec. An input end signal is sent to the pattern editing circuit 8.

Next, the operation of the audio region detection apparatus of the present invention having the above-described configuration will be explained using the input signal waveform diagram shown in FIG. The frequency spectrum of this input signal is extracted by the spectrum extractor 3 from an input signal as shown, which is composed of the audio signal S obtained from the microphone 1 via the microphone amplifier 2 and the ambient noise N.
The data are sequentially extracted at a sunspring period of about 10 msec and stored in the buffer memory 4 in chronological order.
On the other hand, the comparator 5' detects the time point (T ₁ ) when the peak value of the input signal from the microphone amplifier 2 exceeds the threshold value (THR1), and the audio input detection signal is sent to the audio pattern editing circuit 8 for noise pattern editing. The signal is sent to the circuit 11. At this time, the audio pattern editing circuit 8
A time point (T ₀ ) 50 msec earlier than the time point (T ₁ ) is set as the voice input start time point. On the other hand, the noise pattern generation circuit 11 reads out each frequency spectrum of the buffer memory 4 that is included in the time domain before the voice input start time (T ₀ ), that is, in the time domain of only the noise signal, and calculates the time-averaged frequency spectrum. is immediately stored in the noise pattern memory 12 as a noise pattern. Then, the noise region detection circuit 13 detects a coincidence between the noise pattern in the noise pattern memory 12 and the frequency spectrum at each time point sequentially extracted from the spectrum extractor 3 after the time point (T ₁ ), and the audio signal S The noise area detection signal is output at times (T ₂ ′ to T ₃ ′) and after the time (T ₄ ′) when the peak value of the noise N decreases and the ratio of the level of the noise N increases. The counter circuit 7' counts the time during which this noise area detection signal is output, and outputs an audio input end signal when this counted value reaches 150 msec.
The following temporary audio signal interruption times (T ₂ ′ to T ₃ ′)
At this time, this audio input end signal cannot be obtained, and at a time point (T ₅ ′) after 150 msec has elapsed from the time point (T ₄ ′), the audio input end signal is introduced into the audio pattern editing circuit 8, and at this time ( T ₅ ′) is set as the voice input end point. Based on the voice input start time (T ₀ ) and voice input end time (T ₅ ') set in the voice pattern editing circuit 8 in this way, the buffer memory 4 included in the voice area (T ₀ to _{T 5} ') is A frequency spectrum sequence is stored in the voice pattern memory 9.

In the above-mentioned voice area detection device, even when there is no ambient noise, a silent pattern is stored in the noise pattern memory 9, and a silent area is detected by the noise area detection circuit 13. Therefore, the end point of voice input can be definitely set.

In the above explanation, a comparator 5' was shown which obtains an audio input detection signal by comparing the input signal from the microphone amplifier 2 with a certain threshold value (THR1), but instead of this comparator 5', A comparison means may be provided for obtaining an audio input detection signal by comparing the time-series change signal of the frequency spectrum from the spectrum extraction circuit 3 with a certain threshold value. In addition, although the configuration has been shown in which the noise pattern created by the noise pattern editing circuit 11 is stored in the noise pattern memory 12 every time there is an input signal, the speech area detection device according to the present invention is applicable when there is always constant ambient noise. When using a noise pattern, or when a certain amount of hum noise is mixed into the audio signal, it is possible to fix the noise pattern in advance and store it in the noise pattern memory 12.

As is clear from the above description, the audio region detection device of the present invention detects audio input from a comparison circuit that detects when the peak value of the audio waveform or the feature parameter signal waveform from the feature extraction circuit exceeds a certain threshold. A noise parameter extracted from the noise at that time and stored in a noise pattern memory, with a time point a predetermined time before the signal is output as the voice input start point, and with no voice input in advance;
The point in time when the noise region detection signal outputted by the noise region detection circuit when the feature parameters from the feature extraction circuit match are obtained continuously for a specific period of time is set as the voice input end point. Therefore, even if there is voice input in an environment with a high level of ambient noise, it is possible to detect a state in which the voice input is reduced and only noise is input for a specific time.
This eliminates the possibility that a voice interruption point where voice input temporarily stops is mistaken for the end of voice input, and it becomes possible to reliably set the end point of voice input. Furthermore, the present invention is very effective against not only ambient noise but also noise components due to hum noise that is common in this type of device.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a conventional speech region detection device, FIGS. 2a and b are input signal waveform diagrams for explaining the operation of the conventional speech region detection device, and FIG. 3 is a speech region detection device of the present invention. FIG. 4 shows an input signal waveform diagram for explaining the operation of the voice region detection apparatus of the present invention. 1... Microphone, 2... Microphone amplifier,
3... Spectrum extractor, 4... Buffer memory, 5, 5'... Comparator, 7, 7'... Counter circuit, 8... Voice pattern editing circuit, 9... Voice pattern memory, 10... Noise recognition Means, 11...
Noise pattern editing circuit, 12... Noise pattern memory, 13... Noise area detection circuit.

Claims (1)

[Claims] 1. A feature extraction circuit that extracts voice feature parameters from a voice waveform, and a device that indicates that voice input has occurred when the peak value of the voice waveform or the feature parameter signal waveform from the feature extraction circuit exceeds a certain threshold. A comparison circuit that outputs a voice input detection signal, a noise pattern memory that stores a pattern of noise parameters extracted from the noise at that time without any voice input, and feature parameters from the feature extraction circuit. a noise area detection circuit that outputs a noise area detection signal when a match is found with the noise parameters of the noise memory; A voice region detecting device characterized in that a time point at which an input starts is set and a time point at which a noise region detection signal is continuously outputted for a specific time from the noise region detecting circuit is set as a voice input end time point.