JPS5834986B2 - Adaptive voice detection circuit - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voice detection circuit used for DSI (Digital Speech Interpolation) and the like.

In recent years, demand for long-distance communication lines using satellites or submarine cables has rapidly increased.

Since the construction costs of these long-distance lines are high, effective use of telephone lines is required.

TASI (Time As
signment 5peechInterpolat
in hourly allocated audio insertion method), DSI (Digit
al 5peech Interpolation n-
(digital audio insertion method), etc. are known.

These methods all take advantage of the fact that in one call, one speaker spends only about 40% of the time talking.

That is, 60 channels or 120 channels, 2
With a large number of call channels, 40 channels,
By performing processing such as inserting audio from other channels during times when no calls are being made on one channel, calls on channels 60, 120, and 240 can be transferred to channels 30, 60, and 120, respectively. One of the basic methods for these TASI, DSI, etc. is to compress the data into 100% data and transmit it via satellite or submarine cable line.

As can be seen from the above description, in these systems, detection of whether or not audio exists for each channel, that is, audio detection, is fundamentally important.

The characteristics of the voice detection circuit required in these methods are the following two points.

(1) Audio signals should be detectable as quickly as possible.

(2) The probability of erroneously detecting noise as voice is as small as possible.

On the other hand, the level of interline noise varies widely from channel to channel, with some channels being very loud and others being very small.

For these reasons, the threshold value is adaptively changed depending on the noise level of the line, so that when the noise level is high, only large signals are detected, and when the noise level is low, even small signals are detected. Adaptive speech detection circuits configured are known.

In this way, when an adaptive speech detection circuit operates ideally, it has various desirable characteristics, but the circuits that have been realized so far have not been able to measure the noise level, and how to set the threshold value at the measured noise level. There are many unresolved issues in the specific circuit configuration, such as how to control the circuit, and although the circuit is complex, the expected characteristics have not been obtained.

This is because, even though recent systems are capable of sample value processing using the digital method, etc., they still use the conventional analog method, which is not a method that is sufficiently suited to sample value processing.

Therefore, it is an object of the invention to provide an adaptive voice detection circuit with a very simple circuit structure, which is highly suitable for sample value processing, especially digital processing.

The basics of the conventional analog system are as follows.

After rectifying the input signal, it passes through two types of smoothing circuits.

The time constants of these two types of smoothing circuits are selected to be different values, that is, one is selected to be small and the other is large.

Then, using the output value of the smoothing circuit with a larger time constant as a threshold, the output of the other smoothing circuit is checked, and if it is a dog than the threshold, it is determined that there is a voice.

The reason behind this method is that, regardless of the size of the specific number, if the signal is a steady signal of a constant level, the output of both smoothing circuits will take the same value. When the signal level increases rapidly over time, the one with the smaller time constant follows the signal level and increases rapidly, but the one with the larger time constant does not increase rapidly but increases slowly with a delay.

Therefore, by comparing output values of smoothing circuits having different time constants, it is possible to detect the beginning of the audio signal.

This is the principle.

There is no change in principle to the conventional methods using sample values such as digital signals, and only new ideas have been introduced in the level measurement method (for example, Japanese Patent Application No. 47
-124169).

Next, the principle idea of the present invention will be explained using FIG.

In FIG. 1, the solid line indicates the envelope of the input audio waveform.

The basic idea of the present invention is to use a signal shown by a dotted line as an amplitude threshold for voice detection for such an input signal having an envelope.

The signal shown by the dotted line can be obtained by delaying the signal shown by the solid line by T time and further suppressing the maximum value of the signal by M.

It is clear from the figure (on the contrary, at the end of the audio, the input signal falls below the threshold and is not detected, but this is because once the audio signal is detected, the signal becomes smaller and it is determined that there is no audio). There is no problem because it is possible to use a method of setting a so-called hangover time, in which the sound continues for a certain period of about 200 m5 even after the hangover.

In the above explanation, it was explained that the delayed human signal envelope is used as the amplitude threshold for voice detection, but in this case, the signal level measurement value within a short period of time is used instead of the envelope value, and that value is used as the amplitude threshold value for voice detection. It is obvious that a method of determining the threshold value based on the above may also be adopted.

Since the present invention is a delay operation method that is suitable for sample values, particularly digital processing, it can be applied very effectively to systems that use sample values.

FIG. 2 shows an embodiment of the invention.

First, an input signal expressed as a sample value is inputted from an input terminal 100 and sent to a threshold value setting section 110 and a threshold value circuit 101.
be guided by.

The threshold value circuit 101 compares the magnitude of the input sample value with the threshold value controlled by the output of the threshold value setting unit 110, and outputs +1 if the value is above the threshold value, and -1 if it is less than the threshold value.

The output of the threshold circuit 101 is led to the accumulator 102, where the accumulator input values are integrated.

The value accumulated by the accumulator 102 is compared with a predetermined constant value by a subsequent comparison circuit 103, and if the value is greater than or equal to the constant value, it is determined that a voice has been detected, and the information that the voice has been detected is hung over. im circuit 104
Send to.

The hangover time circuit 104 stores information that the voice has been detected, and after receiving the information that the voice has disappeared from the hangover time circuit 103, processes a predetermined period of time as a period with the voice.

On the other hand, in the interval value setting section 110, the level of the input signal is measured by a level measuring circuit 111.

The output of the level measurement circuit 111 has a first storage circuit 112.
A second memory circuit 113 is also connected in cascade, and these memory circuits newly store the respective input values at the time when the pulse (set pulse) signal from the timing pulse input terminal 115 is applied. .

The simplest method for the level measurement circuit 111 is to use, for example, the maximum value within a predetermined period of time.

FIG. 3 shows an embodiment of a circuit for extracting the maximum amplitude value in a short period of time.

First, the amplitude value of the input signal sample (excluding code information) is inputted to the input terminal 1110, and the comparison circuit 1113 compares it with the maximum amplitude value stored in the storage circuit 1112 up to the present time.

The value determined as a dog as a result of the comparison is selected by the selection circuit 1111.
is selected and newly stored in the storage circuit 1112.

A terminal 1114 indicates an input terminal for a timing pulse (clear pulse) signal for clearing the memory contents of the memory circuit.

With such a configuration, the contents stored in the memory circuit 1112 just before a clear pulse is applied are the input signal amplitude values from the point in time indicated by the previous clear pulse to the present moment. This indicates the maximum value of .

Next, FIG. 4 shows the short-time maximum value circuit in the level measuring circuit 11.
An example of a time chart of the threshold value setting unit 110 when used in the case where the threshold value setting unit 110 is used in the example shown in FIG.

First, a indicates the amplitude waveform of the input signal, and in reality, the sample value obtained by sampling this waveform at a fixed period (for example, 125 Ps) is input as described above. It is shown as a continuous waveform in terms of.

b indicates a clear signal for the storage circuit in the maximum value extraction circuit (FIG. 3), and the solid line C indicates the value stored in the storage circuit.

Further, d indicates a set pulse for the memory circuits 112 and 113, and e and f indicate the memory contents of the memory circuits 112 and 113, respectively.

The interval at which the clear pulse is applied is selected to be a relatively large value so as not to cause large variations in the maximum value of line noise (which can be assumed to be Gaussian) within time T.

For example, about 10 to 40 m5 is considered appropriate.

Comparing Figure 4 a and f, we can see that for the input amplitude waveform a, f
It will be understood that if speech detection is performed using a threshold value of , small amplitude parts such as word-initial consonants will be accurately detected.

If T230m5 is selected, referring to the example in Figure 4,
The small amplitude portion at the beginning of a word extends over 30 m5, but in most of that portion, the threshold value is a small value determined from the background noise level, allowing accurate detection.

Also, in the case of a hollow consonant part, etc., the small amplitude part of the consonant part is 150
In some cases, it may continue for about 30 m5, and in this case, according to the above example, the first 30 m5 or so can be detected, but after that,
Since the novel itself of the input consonant serves as a threshold, there is a possibility that the input consonant may not be detected.

However, as mentioned above, it is normal to provide a hangover time in the voice detection circuit, and if a signal of 30m5 or more is detected, it is determined that there is really a voice (not noise) and a hangover time of about 150m5 is provided. Since this is common sense, it turns out that it is not actually a problem.

The memory circuits 112 and 113 have the function of delaying the measured level value by 2T time, but if more delay is required, it is possible to achieve it by cascading more memory circuits. Needless to say.

However, it is important here to connect at least two or more memory circuits in series; using just one memory circuit does not necessarily work.

That is, when two are used, in the example of FIG. 4, the threshold value for the waveform in the section (■) is the measured level in the section (2), which is the two previous section.

Similarly, the threshold values for the intervals ■ and ■ are respectively ■. It is determined by the measurement level of the section (■).

Therefore, as the threshold for the beginning of a speech, the signal level at a time sufficiently distant from that time is used, and this is the most important point.

On the other hand, when there is only one memory circuit, the waveform C in Fig. 4, that is, the signal level of the immediately previous section, for example, the threshold of the section ■ is the section of ■. measurement level is used.

Therefore, there is no problem in a case like the example in FIG. 4, but in a case like in FIG. 5 where the voice starts from the last part of the section 2, the voice is not detected.

That is, in FIG. 5, the audio starts from the section (2).

Since the threshold value for this section is the level of the section (■), it is small, so the input signal amplitude exceeds the threshold value, but in consideration of noise, it is not determined that voice has been detected even if the threshold value is exceeded several times.

In this way, it enters section ■ with no audio detected, but
Since the threshold value for this section is determined by the measurement level of section (2), when performing a simple maximum value extraction type level measurement as shown in Fig. 3, the threshold value is larger than the input signal amplitude as shown in Fig. 5C, Detection becomes impossible.

In such a case, it goes without saying that it is not impossible to use a complicated circuit, such as a smoothing circuit with a sufficiently large time constant, as in the conventional analog system.

On the other hand, the present invention can of course be applied to cases where such a complicated level measurement circuit is used, but as described in the embodiment, it can also be applied to an extremely simple level measurement circuit such as a maximum value extraction circuit. The feature is that it can be applied effectively.

As described above, the present invention is a system that is extremely suitable for sample value or digital processing such as memory circuits, and will greatly contribute to improving the characteristics and reducing costs of DSI, etc., which are expected to become widespread in the future.

[Brief explanation of drawings]

FIG. 1 is a diagram for explaining the present invention in detail. FIG. 2 shows an embodiment of the present invention, in which 101 is a threshold circuit;
102 is an accumulator, 103 is a comparison circuit, 104
110 is a hang-over-quim circuit, 110 is a threshold setting section, 111 is a level measuring circuit, and 112 and 113 are storage circuits, respectively. FIG. 3 is an example of a level measurement circuit using a maximum value extraction circuit, and FIGS. 4 and 5 are time charts for explaining the operation of the example when a level measurement circuit using a maximum value extraction circuit is used. shows.

Claims (1)

[Claims] 1. An adaptive sound detector that operates based on sample values and controls the amplitude threshold according to the background noise level, which is connected in cascade to a level measuring circuit that measures the level of an input signal and the output of the level measuring circuit. a threshold value setting section including a plurality of memory circuits, and a threshold value circuit controlled by the output value of the last stage memory circuit of the plurality of memory circuits, 1. An adaptive voice detection circuit characterized in that the rewriting of the contents of each memory circuit is controlled by a timing pulse given at a cycle longer than the sampling cycle of an input signal.