JP2897628B2 - Voice detector - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech detector which normally performs threshold adaptation even when a speech signal is at a low level and maintains good speech detection characteristics.

[0002]

2. Description of the Related Art For example, a conventional sound detector disclosed in Japanese Patent Application Laid-Open No. 3-141740 has a low level detecting section 1 which detects a low level (a consonant part of a sound, etc.) of a sound input signal 101 as shown in FIG. It compares with the adaptive threshold value 110 from the threshold value setting unit 9 to determine sound / non-sound, and outputs the result as a low-level sound / silence judgment result 102. The high-level detection unit 2 compares a high level (a vowel part or the like of the voice) of the voice input signal 101 with a predetermined threshold value 111 to determine the presence or absence of sound, and outputs the result as a high-level presence / absence determination result 103. I do. The OR operation unit 3 performs an OR operation on the low-level sound / non-sound determination result 102 from the low-level detection unit 1 and the high-level sound / non-sound determination result 103 from the high-level detection unit 2. The sound / silence determination result 104 is output. The hangover adding unit 4 performs a hangover addition process on the sound / non-sound determination result 104 from the logical sum operation unit 3 (a process for holding the sound determination during a predetermined time after the state is changed from the sound to the silent).
To obtain a sound detection output signal 105. The noise level calculator 8a performs an absolute value averaging on samples in the predetermined block section of the audio input signal 101. According to the comparison result 124 between the audio input signal 101 from the high level detection unit 2 and the predetermined threshold value 111, if the audio input signal 101 does not exceed the predetermined threshold value 111 within a predetermined block section, the background noise level 109 is updated. . If it exceeds at least once, the previous calculated background noise level is maintained without updating. The threshold setting unit 9 sets a threshold 110 according to the background noise level 109 from the noise level calculation unit 8a.

The above-mentioned conventional voice detector employs an adaptive threshold type voice detection system which can follow a rapid change in the background noise level.

[0006] As shown in FIG. 12, the low level detector 1 first compares the low level of the audio input signal 101 with the adaptive threshold 110 from the threshold setting unit 9 by the comparator 11. Next, in accordance with the comparison result 125 between the audio input signal 101 and the adaptive threshold value 110, the first determiner 12 determines the presence / absence of sound / no sound for each block of a predetermined time. Further, the second determiner 13 determines that there is sound only when sound is determined for a predetermined number of consecutive blocks, and outputs the result as a low-level sound / silence determination result 102.

As shown in FIG. 13, the high level detector 2 first determines the high level of the audio input signal 101 to a value higher than the adaptive threshold value 110 in advance by the comparator 21 (usually set higher than the maximum background noise level which can be expected). D) Compare with a predetermined threshold value 111. Next, the voice input signal 101 and the predetermined threshold 111
In accordance with the comparison result 124, the first determiner 22 determines the presence / absence of sound / non-speech in units of blocks of a predetermined time. Further, the second determiner 23 determines that there is sound only when the sound is determined for a predetermined number of consecutive blocks, and outputs the result as a high-level sound / non-sound determination result 103. The comparator 21 outputs a comparison result 124 between the audio input signal 101 and the predetermined threshold 111.

[0006] As shown in FIG.
When the average level of the audio input signal 10 is extremely low,
The comparison result 124 between 1 and the predetermined threshold value 111 is shown in FIG.
become that way. Therefore, the high-level detection unit 2 performs a silence determination even for a sound section. The noise level calculator 8a calculates the background noise level 10 not only for silence but also for sounded sections.
9, the background noise level 109 is calculated.
Is higher than the desired background noise level calculation value as shown in FIG. Further, the threshold setting unit 9 increases the adaptive threshold 110 according to the increase of the background noise level calculation value. Next, a comparison result 12 between the voice input signal 101 and the adaptive threshold value 110
5 is as shown in FIG. Therefore, the low-level detection unit 1 performs silence determination for a low-level voice section (a low-voice level section such as the beginning or end of a speech). Also, the logical sum operator 3 outputs the sound / non-sound determination result 104 in the same manner as shown in FIG.

[0007]

The above conventional speech detector employs an adaptive threshold type speech detection system capable of following a sudden change in the background noise level. However, when the speech signal level is extremely low, Since the low-level voice section is also added to the background noise level calculation section and the adaptive threshold is increased, there is a problem that the low-level voice section is determined to be silent.

[0008] Problems to be this invention solves the audio signal is normally threshold adapted even at low levels, is to provide a speech detector to maintain good voice detection characteristic.

[0009]

According to a first aspect of the present invention, there is provided a speech detector comprising a low level detector, a high level detector, a logical sum.
Operation unit, hangover addition unit, signal strength calculation unit, smoothing
Unit, a comparator, a noise level calculator, and a threshold setting unit
A voice detector, wherein the low level detector detects a voice input signal.
Low level sound / no compared to the adaptive threshold from the threshold setting unit
The sound detection result is output, and the high-level detection unit outputs the sound input signal.
Is compared with the predetermined threshold or the adaptive threshold from the threshold setting unit.
Outputs high-level sound / non-speech judgment result and performs logical sum operation
Indicates the low-level sound / silence judgment result and the high-level sound / silence
Performs a logical OR operation on the judgment results to produce a sound / silence judgment result
The hangover addition section adds
Hangover addition processing is performed to
Then, the signal strength calculation unit calculates the signal of the audio input signal for each predetermined section.
The signal strength is calculated and output as a signal strength output .
The signal strength output from the signal strength calculation unit smoothes comparator
Compares the signal strength output with the smoothed output and outputs the comparison result
Then, the noise level calculation unit calculates the signal based on the comparison result of the comparator.
Performs arithmetic averaging on the signal intensity output to obtain the background noise level.
Or the value output immediately before is used as the background noise level.
Output, and the threshold setting unit adapts to the background noise level.
A threshold is calculated and output.

[0010] speech detector according to claim 2 of the present invention, smooth
Section has a smoothing section comparator, up-down counter,
The slip comparator compares the signal strength output and the up / down counter.
By comparing the output to output the result of the comparison, the up-down Cow
Signal strength output based on the comparison result of the smoothing unit comparator.
If it is smaller than the smoothed output, subtract the current value of the counter,
If the signal strength output is greater than the smoothed output,
The present value is added to obtain a smoothed output.

[0011] speech detector according to claim 3 of the present invention, smooth
Unit is a smoothing unit first comparator, an up-down counter, a smoothing unit
Unit second comparator, a selector, the smoothing unit first comparator,
Compare the signal strength output with the output of the up / down counter.
And outputs the comparison result Te, up-down counter, smooth portion
The signal strength output is smoothed according to the comparison result of the first comparator.
If it is smaller, subtract the current value of the counter and output the signal strength.
If the force is greater than the smoothed output, add the current counter value
The smoothing unit second comparator outputs an up-down count.
By comparing the output and the constant A of the motor and outputs a comparison result, selector
Up and down according to the comparison result of the smoothing section second comparator.
Output of counter or constant A as smoothed output
It is.

[0012] speech detector according to claim 4 of the present invention, smooth
The unit includes a first multiplier, a second multiplier, a smoothing unit first comparator,
Smooth part second comparator, smoothing part third comparator, logical sum operator,
Comprising a-down counter, the first multiplier, a constant α and A
Calculate the upper limit by multiplying the output of the up / down counter
And, second multiplier output constant β and the up-down counter
Calculating a lower limit by multiplying the force, the smoothing unit first comparator,
Compare the signal strength output with the output of the up / down counter.
And outputs the comparison result. The second comparator of the smoothing unit outputs the signal strength.
Comparing the output with the output of the first multiplier and outputting a comparison result;
The smoothing unit third comparator outputs the signal strength output and the output of the second multiplier.
And the comparison result is output, and the logical sum
OR of the comparison result between the second comparator and the third comparator
And determine whether the signal strength is within a predetermined range as the determination result.
Output, and the up / down counter is determined by the OR
If the result is within the specified range, the current value is held and the
When the signal is outside, the signal obtained from the comparison result of the smoothing unit first comparator is
When the intensity output is smaller than the output of the up / down counter
Subtracts the current value of the counter and increases the signal strength output.
When the output of the counter is larger than the
The sum is used as a smoothed output.

[0013] Voice detector according to claim 5 of the present invention, smooth
Is a low-pass filter forming a moving average filter
Calculates and outputs a smoothed output from the signal strength output.
Things.

[0014] Voice detector according to claim 6 of the present invention, compared
Even if the result of the comparison changes, the
The result is stored and output.

[0015]

According to the voice detector of the present invention, the voice signal has a low level.
Threshold adaptation normally and maintain good speech detection characteristics
You.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech detector according to one embodiment of the present invention is shown in FIG.
, The low-level detection unit 1, the high-level detection unit 2, the OR operation unit 3, the hangover addition unit 4, and the threshold setting unit 9
This corresponds to FIG. 11 of the above conventional example. The signal strength calculation unit 5
The predetermined number N of samples in the block section of the signal strength calculation period t are subjected to the square-value arithmetic mean with respect to the audio input signal 101, and a signal strength output 106 is obtained. The smoothing unit 6 compares the signal strength output 106 from the signal strength calculation unit 5 with the counter current value at each signal strength calculation cycle t, and adjusts the counter current value according to the result to obtain a smoothed output 107. The comparator 7 compares the signal strength output 106 from the signal strength calculation unit 5 with the smoothed output 107 from the smoothing unit 6, and when the signal strength output 106 is larger than the smoothed output 107, 1 (ON);
When smaller, 0 (off) is output as the comparison result 108. The noise level calculator 8 calculates the comparison result 1 from the comparator 7.
When 08 is the 0 section, the signal strength output 106 from the signal strength calculation unit 5 is arithmetically averaged to obtain a background noise level 10
Updated as 9. At the time of one section, the immediately preceding background noise level calculation value is maintained without updating.

The voice detector of the above-described embodiment employs a threshold adaptive voice detection system that performs threshold adaptation normally even when a voice signal is at a low level and maintains good voice detection characteristics.

As shown in FIG. 2, the signal strength calculator 5
The voice input signal 101 is sampled by the multiplier 51 every sampling period.
Is calculated. Next, for a predetermined number N of samples in the block section of the signal strength calculation cycle t, the cumulative adder 52 resets the output value to 0 at the time of initial setting (immediately after power-on) and immediately after output latching. And the accumulator 52
And the current value of. Further, a multiplier 54 multiplies the output of the latch 53 of the added value by a constant 1 / N, and performs an arithmetic mean of square values to obtain a signal strength output 106.

As shown in FIG. 3, the smoothing unit 6 compares the signal strength output 106 from the signal strength calculation unit 5 with the current value of the up / down counter 62 by the comparator 61 at every signal strength calculation period t. 1 when intensity output 106 is greater
(On), when it is small, 0 (Off) is output as the comparison result 113. Next, when the comparison result 113 from the comparator 61 in the up / down counter 62 is 1 or 0, the counter current value is added or subtracted by 1 to obtain a smoothed output 107 to the comparator 7.

When the average level of the audio input signal 101 is extremely low as shown in FIG. 4A, first, when the audio input signal 101 changes state from silence to speech as shown in FIG. The signal strength output 106 of the calculator 5 increases and becomes larger than the smoothed output 107. FIG.
As shown in (c), the signal strength output 106 and the counter current value 1
07 as 1 and the signal strength output 106
The smoothed output 107 continues to be increased by 1 until it becomes equal to When the state changes from sound to silence, signal strength output 1
06 decreases and becomes smaller than the smoothed output 107. FIG.
As shown in (c), the signal strength output 106 and the counter current value 1
07 and the signal strength output 106
If the signal intensity calculation cycle t (update cycle of the smoothed output 107) is set to be sufficiently long, the smoothed output 107 is changed to a state of sound or no sound. The change can be made sufficiently gently as compared with the accompanying change in the audio signal strength, and the smoothed output 107 can be set to a value between the sound signal intensity at the time of sound and at the time of no sound. Further, the comparator 7 compares the signal strength output 106 with the smoothed output 107, and roughly determines whether the audio input signal 101 is in a sound state or in a silent state. When the comparison result between the signal strength output 106 and the smoothed output 107 from the comparator 7 is in the 0 section (the section in which the voice input signal 101 is determined to be in a silent state), the noise level calculation section 8 calculates the background noise level 109. calculate. Addition of a sound section to the noise level calculation section is reduced, and the error in the background noise level calculation value can be reduced even when the audio signal level is low as shown in FIG. Next, the threshold setting unit 9 adjusts the noise level calculated value to the threshold value 110.
Therefore, as shown in FIG. 4A, the adaptive threshold value 110 is a sufficiently low value to determine whether the speech head or the tail part having a low audio signal level is sound. Therefore, in the low-level sound / non-speech determination result 102, a sound period is correctly determined as a sound as shown in FIG. Low level sound / silence judgment result 1
02 and high-level sound / silence determination result 103 (FIG. 4D)
4), a sound / silence determination result 104 in which an audio signal is correctly detected is obtained as shown in FIG.

In the above embodiment, as shown in FIG. 5, the smoothing section 6 is provided with a second comparator 63 and a selector 64 to form a smoothing section 6a . The second comparator 63 compares it with a predetermined constant A, and outputs 1 (ON) when the smoothed output 107 is larger, and outputs 0 (OFF) as the comparison result 115 when the smoothed output 107 is smaller. Next, when the comparison result 115 from the second comparator 63 is 1 or 0 in the selector 64, the constant A or the smoothed output 107 may be selected as the selected smoothed output 107a to the comparator 7. Signal with long sound duration (F
AX transmission / reception modem signal, etc.), the smoothed output 107 continues to increase during a sound, so that a sound section after the smoothed output 107 matches the signal strength output 106 is also included in the background noise level calculation section, and the background noise level is calculated. This has the effect of solving the problem of causing an error in the noise level calculation value. When the voice input signal 101 has a long sound duration as shown in FIG.
When the state of the audio input signal 101 changes from silence to sound as shown in FIG.
Is larger than the smoothed output 107. The smoothing part 6a
As shown in FIG. 6c, the signal strength output 105 and the counter current value 1
The comparison result 113 with 07 is set to 1, and the smoothed output 107 starts to be increased by one. FIG. 6 shows whether the constant A is exceeded or not.
As shown in FIG. 6D, the comparison result 115 between the smoothed output 107 and the constant A is set to 0 or 1, and as the selected smoothed output 107a, as shown in FIG.
Select Even when the sound duration is long, the constant A becomes the upper limit value, and the selected smoothed output 107a becomes the signal intensity output 1
06 will not be exceeded.

In the above embodiment, as shown in FIG. 7, the smoothing unit 6 includes first, second and third comparators 61a, 61b and 61c, an up-down unit 62a, and a logical sum operation unit 65.
And a smoothing unit 6b including first and second multipliers 66 and 67, which has the same effect as in FIG.
At each signal strength calculation period t, first, the signal strength output 106 from the signal strength calculation unit 5 is output to the first, second, and third comparators 6.
At 1a, 61b and 61c, the counter current value is 10 respectively.
7b, a counter current value 107b and a predetermined constant α (α
> 1) and the current counter value 107
multiplication result 11 of b and a predetermined constant β (0 <β <1)
Compare with 9. The first comparator 61a outputs the signal strength output 1
When 06 is larger than the current counter value 107b, 1 (ON) is output, and when 0 is smaller, 0 (OFF) is output as a comparison result 120 between the signal strength output 106 and the current counter value 107b. The second comparator 61b determines 1 (on) when the signal strength output 106 is larger than the first multiplication result 118 , and 0 (off) when the signal strength output 106 is smaller than the first multiplication result 118. The comparison result between the signal strength output 106 and the first multiplication result 118 Output as 121. The third comparator 61c determines 0 (off) when the signal strength output 106 is larger than the second multiplication result, and 1 (on) when the signal strength output 106 is smaller than the second multiplication result. The comparison result 122 between the signal strength output 106 and the second multiplication result 119 Output as Next, the second and third comparators 61
The logical sum operation unit 65 performs a logical sum operation on the comparison results 121 and 122 from b and 61c, and outputs the result as a signal strength determination result 123. Up / down counter 6
2a, the signal strength determination result 12 from the logical sum operator 65
When 3 is 0, the current counter value 107b is held as the previous value. In the case of 1, the comparison result 12 from the first comparator 61a
When 0 is 1, the counter current value 107b is incremented by 1; when it is 0, the counter current value 107b is decremented by 1 to obtain a decision smoothed output 107b. When the voice input signal 101 has a long sound duration as shown in FIG.
When the state of the audio input signal 101 changes from silence to sound as shown in (b), the signal intensity output 10
6 is larger than the judgment smoothed output 107b. The smoothing unit 6b compares the signal strength output 106 with the current counter value 107b 120 and compares the signal strength output 106 with the first multiplication result 118 as shown in FIGS. 8C, 8D and 8F. The result 121 and the signal strength determination result 123 are each set to 1, and the determination smoothed output 107b starts to increase by one. When the signal strength output 106 becomes equal to the first multiplication result 118, the comparison result 121, the comparison result 122, and the judgment result 123 are set to 0 as shown in FIGS. 8D, 8E, and 8F, and the judgment smoothing is performed. 8 as the output 107b.
As shown in (b), the previous value is maintained. When the state changes from sound to silence, the signal strength output 10
6 is smaller than the decision smoothing output 107b. As shown in FIGS. 8C, 8E, and 8F, the smoothing unit 6b compares the signal strength output 106 with the counter current value 107b.
20, the comparison result 122 between the signal strength output 106 and the second multiplication result 119 and the signal strength determination result 123 are set to 0, 1, and 1, respectively, and the determination smoothed output 107b starts to decrease by one. When the signal strength output 106 is equal to the second multiplication result 119, the comparison result 121, the comparison result 122, and the judgment result 123 are set to 0 as shown in FIGS. 8D, 8E, and 8F, and the judgment smoothing is performed. Output 10
As shown in FIG. 8B, the previous value is maintained as 7b. Even when the sound duration is long, the first multiplication result 118 becomes the upper limit value and the second multiplication result 119 becomes the lower limit value, and the judgment smoothed output 107b takes a value between the sound intensity and the signal intensity at the time of no sound. Is controlled.

In the above embodiment, the first to third adders 600, 601 and 602, as shown in FIG.
To 4 delay elements 611, 612, 621, and 622 and first to fifth multipliers 630, 631, 632, 641, and 6
A low-pass filter comprising 42 may be provided. There is the same effect as in FIG. In FIG. 9, the sampling interval is set to the signal strength calculation cycle t, time (k × t, k = 0, ± 1, ± 2,
If the input and output signal when ...) and X _k and Y _k, if
For example, if b _n = 0, a moving average filter is formed as follows. To: (number of taps n = 0~N N) = 1 / N + 1 Σ n X kn (a n = 1 / N + 1 if) the signal intensity output 106 as an input signal _{X k Y k = Σ n a} n X kn since performing a filtering operation in the block section of the signal strength calculation cycle t, the low-pass smoothing output as an output signal Y _k 1
07c is the squared arithmetic mean of the input samples in the block section of the moving average time width (N + 1) × t. If the moving average time width (N + 1) × t a signal strength calculation block time setting the N to be sufficiently wider than the width t, also include a voiced / silent to the moving average time in bandwidth, low-pass smoothing output 107
c takes a value between the sound level at the time of speech and the background noise level at the time of silence.

In the above-described embodiment, the output holding means is provided in the comparator 7 to form a comparator 7d. When the comparison result 108 of the signal strength output 106 and the smoothed output 107 changes from 1 to 0, the predetermined time t1 is reduced. Control may be performed such that the output is held at 1 until the time elapses and the output is held at 0 until the state changes from 0 to 1. This has the effect of preventing the noise level calculation unit 8 from adding a low-level voice section (such as an end part) to the background noise level calculation section. When the average level of the audio input signal 101 is extremely low as shown in FIG. 10A, first, when the state of the audio input signal 101 changes from silence to speech as shown in FIG. Signal strength output 106 increases and becomes larger than the smoothed output 107. The smoothing unit 6 keeps increasing the smoothed output 107 by one until it becomes equal to the signal strength output 106. Comparator 7
Sets the comparison result 108 between the signal strength output 106 and the smoothed output 107 to 1 as shown in FIG. When the state changes from the voiced to the silence, the signal strength output 106 is reduced small,
It becomes smaller than the smoothed output 107. Signal strength output 106
The smoothed output 107 is continuously reduced by 1 until the value becomes equal to. The comparison result 108 between the signal strength output 106 and the smoothed output 107 is set to 0. As shown in FIG.
If 08 is a state change from 1 to 0, until when a predetermined time elapses t _1, the comparison result 108a of the signal strength output 106 and the smooth output 107 after holding additional function as shown in FIG. 10 (d) The comparison result 108a is set to 1 until the state of the comparison result 108 changes from 0 to 1. Therefore, assuming that the 0 section of the comparison result 108 is the background noise level calculation section as it is as shown in FIG. 10C, a low-level speech section is also added to the background noise level calculation section. As described above, this can be prevented by setting the 0 section of the comparison result 108a as the background noise level calculation section.

In the above embodiment, the high level detector 2 may use a threshold adapted to the background noise level in the threshold setting unit 9 instead of the predetermined threshold 111. In addition, the signal strength calculator 5
The signal strength output 106 may use an absolute value average value or a peak value instead of the square value average value. Further, the smoothing unit 6 may increase / decrease the smoothed output 107 by a fixed amount instead of by one. When increasing / decreasing, a different value may be used instead of a fixed amount.

As described above, in this embodiment, first,
Adaptive threshold for low and high levels of input signal
And a predetermined threshold value, and discuss the sound / non-speech judgment result.
Performs a logical sum operation and hangover addition processing to output a voice detection output signal.
No. Then the signal strength calculation cycle Bro audio input signal
For a given number of samples in the
Comparison result between the signal strength output to be averaged and the current counter value
To add or subtract the current value of the counter to obtain a smoothed output. Or flat
The smoothed output is calculated based on the result of comparing the smoothed output with a predetermined constant.
Select one of constants and output. Or the signal
Intensity output, counter current value, first multiplication result (counter
Multiplication result of current value and first predetermined constant) and
Multiplication result of 2 (counter current value and second constant determined in advance
And the first, second and third comparison results with
Accordingly, the signal strength for performing a logical sum operation on the second and third comparison results
The counter is determined by the combination of the degree judgment result and the first comparison result.
Adds, subtracts, or retains the previous value and outputs it. Or signal strength calculation
A filter for the signal strength output input every output cycle
The ring operation is performed and the signal is transmitted within the block section of the moving average time width.
The signal intensity is smoothed and output. Further, the signal strength output and
Output is held as a result of comparison with the smoothed output or
Control and maintain the signal strength output by arithmetic averaging.
Background noise level to keep the previous value without updating
And set a threshold.

[0027]

As described above, in the speech detector according to the present invention, the noise level calculation unit performs the processing for each predetermined section of the speech input signal.
, Calculate the signal strength, and this signal strength is input first,
Compared with the signal strength of the smoothed audio input signal,
Determines the sound level and an adaptive threshold adapted to this background noise level
Since the threshold setting unit calculates and outputs the value, the audio signal is low level.
Prevents the noise level calculation value from rising accidentally even at the bell
Threshold adaptation can be performed normally, and good voice detection characteristics can be maintained.

[Brief description of the drawings]

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

FIG. 2 is a functional block diagram of a signal strength calculator shown in FIG. 1;

FIG. 3 is a functional block diagram of a smoothing unit shown in FIG. 1;

FIG. 4 is a view for explaining the operation of the voice detector shown in FIG. 1;

FIG. 5 is a functional block diagram of another embodiment of the smoothing unit shown in FIG. 1;

FIG. 6 is a view for explaining the operation of the smoothing unit shown in FIG. 5;

FIG. 7 is a functional block diagram of another embodiment of the smoothing unit shown in FIG. 1;

FIG. 8 is a view for explaining the operation of the smoothing unit shown in FIG. 7;

FIG. 9 is a diagram illustrating a low-pass filter used as a smoothing unit shown in FIG. 1;

FIG. 10 is a view for explaining an output holding function added to the comparator shown in FIG. 1;

FIG. 11 is a functional block diagram of a conventional voice detector.

FIG. 12 is a functional block diagram of a low-level detection unit shown in FIG. 11;

FIG. 13 is a functional block diagram of a high-level detection unit shown in FIG. 11;

FIG. 14 is a view for explaining the operation of the voice detector shown in FIG. 11;

[Explanation of symbols]

1 low level detection section, 2 high level detection section, 3 OR operation section, 4 hangover addition section, 5 signal strength calculation section, 6 smoothing section, 7 comparator, 8 noise level calculation section,
9 Threshold setting unit, 101 voice input signal, 102 low-level voice / non-voice determination result, 103 high-level voice / non-voice determination result, 104 voice / non-voice determination result, 105 voice detection output signal, 106 signal strength output, 107 Smoothed output, 108 comparison result between signal strength output and smoothed output, 109 background noise level, 110 adaptive threshold, 111
Predetermined threshold, 112 constant 1 / N.

Continuation of the front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 G10L 9/00

Claims (6)

(57) [Claims] 1. Low level detector, high level detector, logic
Sum calculation unit, hangover addition unit, signal strength calculation unit, flat
Smooth section, comparator, noise level calculation section, threshold setting section
The low-level detector detects an audio input signal from a threshold setting unit.
Outputs low-level sound / no-sound judgment result in comparison with response threshold
The high-level detection unit detects the audio input signal as a predetermined threshold or a threshold.
High level sound / silence judgment compared to the adaptive threshold from the setting unit
The logical sum operation unit outputs a low level sound / non-speech judgment result and a high level
Performs a logical OR operation on the voiced / silent judgment result to generate voiced / silent
The judgment result is output, and the hangover adding unit hangs on the sound / silence judgment result.
An over-addition process is performed to generate a voice detection output signal, and the signal strength calculation unit determines the signal strength of the voice input signal for each predetermined section.
Calculates the degree and outputs it as a signal strength output, and the smoothing unit smoothes the signal strength output from the signal strength calculation unit
The comparator compares the signal strength output with the smoothed output and
And the noise level calculator calculates the signal strength based on the comparison result of the comparator.
When the output is smaller than the smoothed output,
Apply arithmetic averaging and output as background noise level, signal strength
If the degree output is larger than the smoothed output, the value output immediately before
Is output as a background noise level, and the threshold setting unit calculates an adaptive threshold value adapted to the background noise level.
Output sound detector. 2. The smoothing section includes a smoothing section comparator, and an up-down converter.
It has a counter, a smoothing unit comparator, a signal strength output and an up-down counter.
The up / down counter outputs the result of comparison by the smoothing unit comparator.
If the signal strength output is smaller than the smoothed output,
Subtract current value and signal strength output is greater than smoothed output
2. The voice detector according to claim 1 , wherein at the time, the current value of the counter is added to obtain a smoothed output . 3. The smoothing unit includes a smoothing unit first comparator and an up-converter.
A smoothing unit second comparator and a selector. The smoothing unit first comparator has a signal strength output and an up / down counter.
The up / down counter outputs the comparison result of the smoothing unit first comparator.
When the signal strength output is smaller than the smoothed output,
Subtracts the current signal value and the signal strength output is greater than the smoothed output.
When the threshold is high, the counter current value is added and output, and the second comparator of the smoothing unit determines the output of the up / down counter as constant.
The number A is compared and the comparison result is output, and the selector is increased according to the comparison result of the smoothing unit second comparator.
The output of the down counter or the constant A is used as the smoothed output.
Speech detector according to claim 1, wherein that. 4. The smoothing unit includes a first multiplier, a second multiplier, and a flat multiplier.
Smooth part first comparator, smoothing part second comparator, smoothing part third comparison
, An OR operation unit, and an up / down counter, and the first multiplier has a constant α and an output of the up / down counter.
And the second multiplier calculates the upper limit by multiplying the constant β by the output of the up / down counter.
To calculate the lower limit, and the smoothing unit first comparator compares the signal strength output with the up / down count.
The smoothing unit second comparator compares the signal strength output with the output of the first multiplier.
And outputs a comparison result, and the smoothing unit third comparator outputs the signal strength output and the output of the second multiplier.
And a comparison result is output, and the OR operation unit includes a smoothing unit second comparator and a smoothing unit third comparator
OR operation of the comparison result of
Is output as a determination result, and the up / down counter indicates that the determination result of the OR
When the current value is within the specified range, the current value is held.
Is a signal strength output based on the comparison result of the smoothing unit first comparator.
Is smaller than the output of the up / down counter,
The current value is subtracted, and the signal strength output is
If the output of the counter is larger than the
2. The sound detector according to claim 1, wherein the sound output is a smoothed output . 5. The smoothing unit forms a moving average filter.
Equipped with low-pass filter, smoothed output from signal strength output
The voice detector according to claim 1, wherein the voice detector calculates and outputs the following . 6. The comparator according to claim 1 , wherein said comparison result changes.
4. The method according to claim 1, wherein the comparison result immediately before the fixed time is held and output.
6. The voice detector according to any one of 5 .