JPH0240700A - Voice detecting device - Google Patents

Abstract

PURPOSE:To detect a voice section without reference to variation in voicing level by finding the 1st and 2nd means amplitude values of an input signal and the zerocross frequencies of a voice signal and a reference frequency, and comparing the zerocross frequencies while monitoring the mean amplitudes. CONSTITUTION:When the output value of a long-time mean amplitude calculating circuit 12 is larger than the output value of a short-time mean amplitude calculating circuit 11, a comparator 15 outputs a high-level signal and the output of an OR circuit 18 is a high-level signal. When the counted value of a zerocross counting circuit 13 increases above the counted value of a reference frequency zerocross counting circuit 14, a comparing circuit 16 outputs a high-level signal and the output of the OR circuit 18 is a high-level signal as well. This phenomenon is used to find the 1st mean amplitude value when the rising time constant of a low-pass filter is <=1 millisecond and the falling time constant is 70-100 milliseconds and the 2nd mean amplitude value when the rising time constant is 2-10 seconds and the falling time constant is 1-2 seconds, thereby outputting a detection signal when the zerocross counted value is larger than a reference value.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application This invention relates to a voice detection device used in a voice recording device or the like.

7. Prior Art Conventionally, this type of voice detection method includes, for example,
As described in Japanese Patent No. 65527, a method is known in which an input signal is compared with a set fixed reference value, and when the input signal exceeds the reference value, it is determined as a voice section.

FIG. 3 is a block diagram showing a conventional voice detection device.

In FIG. 3, 1 is an average amplitude calculation circuit that calculates the average amplitude of the input signal, 2 is a reference amplitude setting circuit that sets a reference amplitude value, and 3 is a comparator, which is the average amplitude value that the average amplitude calculation circuit 1 outputs. and the reference amplitude value outputted by the reference amplitude setting circuit 2.

Next, the operation of the above conventional example will be explained.

In the above conventional example, the comparator 3 compares the average amplitude value of the average amplitude calculation circuit 1, which calculates the average amplitude value (or effective value) of the input signal at regular intervals, with the reference amplitude value set in the reference amplitude setting circuit 2. An input signal having an average amplitude value exceeding the reference amplitude value is determined to be voice.

In addition to the above-mentioned conventional examples, detection accuracy may be improved by setting the reference amplitude value in two stages, or a non-speech section for a certain period of time before and after a speech section is forcibly regarded as a speech section by calling it a hangover. Speech detection devices that prevent speech from being cut off at the beginning or end of speech, and speech detection devices that detect speech sections from the zero-crossing frequency of input signals below a fixed reference amplitude value and prevent erroneous deletion of speech sections have also been proposed. .

Problems to be Solved by the Invention However, with the conventional voice detection device described above, when the average amplitude value of the entire input voice is small, erroneous voice deletion tends to occur, and when background noise is large, it is difficult to distinguish between voice and noise. This invention solves the problem of the prior art, and is capable of stably detecting voice sections without being affected by changes in voice level or background noise level. It is an object of the present invention to provide a voice detection device that can perform voice detection.

Means for Solving the Problems In order to achieve the above object, the present invention provides a low-pass filter whose rise time constant is 1 millisecond or less and whose fall time constant is 70 milliseconds to 100 milliseconds. In addition to determining the average amplitude value, the second average amplitude value of the input signal is determined with the rise time constant of the low-pass filter being 2 seconds to 10 seconds and the fall time constant being 1 second to 2 seconds. In addition to counting the frequency, the zero-crossing frequency of the reference frequency is determined, and when the second average amplitude value is greater than or equal to the first average amplitude value, a voice detection signal is output, and the zero-crossing count value of the audio signal is the zero-crossing of the reference frequency. A voice detection signal is output when the count value is greater than or equal to the count value.

Effects The present invention has the following effects due to the above structure. The average amplitude of the input signal over a relatively long time varies relatively slowly depending on the background noise level, and a sudden change in amplitude during voice input appears in the average amplitude over a relatively short time, regardless of the vocalization level. Therefore, the average amplitude over a relatively short period of time is monitored using this average amplitude over a relatively long period of time as a reference, and the voice section is detected. Unvoiced sounds whose amplitude level is still small and are difficult to distinguish from noise and amplitude alone have a higher zero-crossing frequency than ambient noise, and are detected by comparing the zero-crossing frequency with the zero-crossing frequency that coincides with the reference frequency.

Therefore, according to the present invention, there is an effect that a voice section can be detected regardless of fluctuations in vocalization level or fluctuations in ambient noise.

Embodiment FIG. 1 is a block diagram showing a voice detector according to an embodiment of the present invention.

In FIG. 1, 11 is a short-time average amplitude calculation circuit that calculates a relatively short-term average amplitude value of an input audio signal, and 12 is a long-term average amplitude calculation circuit that calculates a relatively long-term average amplitude value of an input audio signal. The calculation circuit includes a zero-crossing counting circuit 13 that counts the zero-crossing frequency of the input audio signal, and a reference zero-crossing counting circuit 14 that counts the zero-crossing frequency of the reference frequency.

Note that each of the average amplitude calculation circuits 11 and 12 passes the full-wave rectified output through a low-pass filter to obtain a desired average amplitude value.

Reference numeral 15 indicates a comparator, the output signal (short-time average amplitude value) of the short-time average amplitude calculation circuit 11 is supplied to the inverting terminal, and the output signal (long-time average amplitude value) of the long-term average amplitude calculation circuit 12 is supplied to the non-inverting terminal. value), and outputs a high-level voice detection signal when the long-term average amplitude value is greater than or equal to the short-term average amplitude value.

Reference numeral 16 indicates a zero-cross comparison circuit, which compares the output signal (zero-cross count value) of the zero-cross counting circuit 13 and the output signal (reference frequency zero-cross count value) of the reference frequency zero-cross counting circuit 14, and determines that the zero-cross count value is equal to the reference frequency zero-cross count value. It outputs a high-level audio detection signal when the value is greater than or equal to the numerical value. 0 17 is an AND circuit that receives the output signal of the comparator 150 inverted output color zero cross comparison circuit 16, and 18 is the output signal of the comparator 15. An OR circuit is shown in which the output signal of the AND circuit 16 and the output signal of the AND circuit 16 are input.

Next, the operation of the above embodiment will be explained.

First, the long-term average amplitude value of the input signal fluctuates relatively slowly depending on the background noise level, and rapid amplitude changes during audio input appear as a short-term average amplitude value regardless of the generation level. If the long-term average amplitude value output by the amplitude calculation circuit 12 is greater than or equal to the short-term average amplitude value output by the short-term average amplitude calculation circuit 11, a high-level signal is output from the comparator 15, so the OR circuit A high level signal is obtained at the output of 18.

In the case of non-speech, which has a small amplitude level and is difficult to distinguish from noise and amplitude alone, the zero-cross count value of the zero-cross counting circuit 13 is larger than the reference frequency zero-cross count value of the reference frequency zero-cross counting circuit 14, so the zero-cross count value is the standard. If the frequency is equal to or higher than the zero-cross count value, a high-level signal is output from the zero-cross comparator circuit 16, so that a high-level signal is obtained at the output of the OR circuit 18.

However, if the time constant of the low-pass filter in the average amplitude calculation circuits 11 and 12 is not appropriate, inconveniences may occur, such as sound interruptions at the beginning or end of a word, or conversely, a silent section being added more than necessary after the end of the speech. .

Therefore, as shown below, both time average amplitude calculation circuit 11
．． The time constants of each of the 12 low-pass filters were determined.

First, in the low-pass filter of the short-time average amplitude calculation circuit 11, the rising time is set to substantially within 1 millisecond using a diode so that the sound can catch up with the steep rising sound.

This is because if you look at the voice pitch waveform, the frequency of the waveform with a large peak is around 100Hz to 300Hz, but if you look at the constituent elements, the main component is a frequency around IKHz. This is because a rise time constant of 1 millisecond or less is required to follow a frequency of about 100 ms.

If the falling time constant is not faster than the long-term average amplitude value, silence after the end of the voice will be mistaken for voice, and if it is too fast, it will cause sound to break.

Therefore, an appropriate time constant was determined by computer simulation as described below.

Next, in the low-pass filter of the long-term average amplitude calculation circuit 12, the condition is that the 7th rise cannot catch up with the short-term average amplitude value in the voice section, and the expiration stage of the voice (-
Since the average duration (time taken to speak with a breath) is about 2 seconds, the long-term average amplitude value needs to rise over 2 seconds.

However, if the start-up is too slow, it will not be able to catch up with the ambient noise level, and the silent section will be mistaken for a voice section.

Therefore, in a computer simulation, even if the rise time constant of the short-time average amplitude calculation circuit 11 was extended to about 100 seconds, there was no change in the compression ratio (the experiment was performed with an S/N ratio of 7 dBt). ~io seconds.

The falling time constant was set to 1 to 2 seconds so that the average level of ambient noise was reached during this period, since the pause between exhalation stages (pause such as breathing) during voice generation was about 1 second.

FIGS. 2(a) to 2(c) are waveform diagrams for explaining the falling time constant of the short-time average amplitude calculation circuit.

In Fig. 2, IIs I2 and 2 are input signals (audio signals), Sl, S2, and S2 are short-time average amplitude outputs, and Ll
%L2, L2 indicates the long-term average amplitude output.

As is clear from Figure 2, the falling time constant is 100.
It can be seen that if it is less than milliseconds, the short-time average amplitude value cannot catch up with the long-term average amplitude value, and silence cannot be detected.

On the other hand, it can be seen that if the duration is not longer than 70 milliseconds, the ending of the voice may be cut off.

Therefore, it is appropriate that the falling time constant of the low-pass filter of the short-time average amplitude calculation circuit 11 is between 70 milliseconds and 100 milliseconds.

In the above embodiment, the comparator 15, the zero-cross comparison circuit 16, the AND circuit 17, and the OR circuit 18 constitute the comparison means, but other configurations that operate in the same manner may be used.

The short- and long-term average amplitude calculating means may have a configuration in which the full-wave rectifier is shared.

Furthermore, a margin of about several decibels is provided to prevent the short-time average amplitude calculation means from erroneously determining the rise due to noise as voice.

Effects of the Invention As is clear from the above embodiments, the present invention provides a first average amplitude of an input signal in which the rise time constant of the low-pass filter is 1 millisecond or less and the fall time constant is 70 milliseconds to 100 milliseconds. In addition to finding the value, find the second average amplitude value of the input signal with the rising time constant of the low-pass filter being 2 seconds to 10 seconds and the falling time constant being 1 second to 2 seconds, and also find the zero-crossing frequency of the audio signal. At the same time, the zero-cross frequency of the reference frequency is determined, and when the second average amplitude value is greater than or equal to the first average amplitude value, an audio detection signal is output, and the zero-cross count value of the audio signal is greater than or equal to the zero-cross count of the reference frequency. Since the voice detection signal is output at this time, there is an effect that the voice section can be stably detected without being affected by changes in the voice level or background noise level.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a voice detection device according to an embodiment of the present invention, FIGS. FIG. 3 is a block diagram showing a conventional voice detection device. 11... Short-time average amplitude calculation circuit, 12... Long-term average amplitude calculation circuit, 13... Zero cross counting circuit, 14
...Reference frequency zero cross counting circuit, 15... Comparator, 16... Zero cross comparison circuit, 17... AND
Circuit, 18...OR circuit. Name of agent: Patent attorney Shigetaka Awano and one other person (Figure 2), 5-・yoko-cane 15mm/7-・ANDday), 131. , OR each nol

Claims (1)

[Claims] Set the rise time constant of the low-pass filter to 1 millisecond or less.
a first calculation means for calculating the average amplitude value of an input signal with a falling time constant of 70 milliseconds to 100 milliseconds, and a low-pass filter with a rising time constant of 2 seconds to 10 seconds and a falling time constant of 1 second or more; a second calculating means for calculating the average amplitude value of the audio signal for 2 seconds; a first counting means for counting the zero-crossing frequency of the input signal; a second counting means for counting the zero-crossing frequency of the reference frequency; outputs a voice detection signal when the average amplitude value of the second calculation means is greater than or equal to the average amplitude value of the first calculation means; A voice detection device comprising means for outputting a voice detection signal when the value is greater than or equal to a numerical value.