JP2004102161A - Device, method, and program for voice detection - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To suppress deterioration in precision of speaking detection due to influence of a speaking time. <P>SOLUTION: A speaker signal is inputted to a short-time sound pressure averaging unit 4a after being passed through high-pass filters 1 and 2a to limit the band almost to 2,000 Hz at which formant frequencies of a human being are mainly distributed, and the speaker signal is inputted to a long-time sound pressure averaging unit 4b after being passed through the high-pass filter 1 and a low-pass filter 2b to limit the band to a band wherein low-frequency noise is distributed. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a voice detection device, a voice detection method, and a voice detection program, and is particularly suitable when applied to a case where detection of a speaker's utterance is accurately performed.
[0002]
[Prior art]
In a conventional voice detection device, there is a method of measuring the sound pressure of a speaker signal sent from the speaker side and comparing the measured sound pressure with a preset threshold value in order to determine whether the speaker is speaking or not.
Further, in order to improve the determination accuracy of a speaker's utterance and no utterance in a noisy environment, for example, as disclosed in Patent Document 1, time averagers having different time constants are provided, and a short-time average sound is provided. There is also a method of subtracting the average sound pressure for a long time from the pressure and comparing the result with a preset threshold value.
[0003]
FIG. 3 is a block diagram showing a schematic configuration of a conventional voice detection device.
In FIG. 3, a conventional voice detection device includes an absolute value conversion circuit 11 for converting a speaker signal into an absolute value, a short-time sound pressure averager 12a for calculating a short-time average sound pressure, and a long-time average sound pressure. A subtractor 13 for subtracting the output of the long-time sound pressure averager 12b from the output of the long-time sound pressure averager 12b and the short-time sound pressure averager 12a to be calculated, and a comparison of comparing the output from the subtractor 13 with a threshold value A vessel 14 is provided.
[0004]
Here, the speaker signal includes a voice signal generated by the speaker's speech and a noise signal in a noise environment.
When the speaker signal is input to the absolute value conversion circuit 11, the speaker signal is converted to an absolute value by the absolute value conversion circuit 11 and output to the short-time sound pressure averager 12a and the long-time sound pressure averager 12b. Is done.
[0005]
The short-term average sound pressure of the absolute-valued speaker signal is calculated by the short-time sound pressure averager 12a, and the long-time average sound pressure of the absolute-valued speaker signal is calculated for a long time. The sound pressure averaging device 12b calculates the sound pressure, and the average sound pressure is subtracted from each other by the subtractor 13.
Here, the noise occurs almost constantly, while the speaker speaks intermittently.
[0006]
For this reason, the average sound pressure of the noise maintains a substantially constant value without depending on the averaging time, and the average sound pressure of the noise captured by the short-time sound pressure averaging device 12a and the long-time sound pressure averaging device 12b substantially coincides with each other. Therefore, the influence of noise can be canceled by subtracting the output value of the long-time sound pressure averager 12b from the output value of the short-time sound pressure averager 12a. On the other hand, when the average time is long, the average sound pressure of the audio signal averages the level during utterance and the level during no utterance, but when the average time is short, the average sound pressure fluctuates according to the speaker's utterance or no utterance. It is possible to extract the level when the speaker speaks or when the speaker does not speak.
[0007]
As a result, in the short-time sound pressure averaging device 12a, it is possible to make the average sound pressure greatly different between the time of utterance and the time of no utterance, and it is possible to extract the level when the speaker utters or when there is no utterance. In addition, the long-time sound pressure averaging device 12b can maintain the average sound pressure substantially constant regardless of utterance or no utterance, and can extract the average sound pressure of noise.
[0008]
Then, the output from the subtractor 13 is input to the comparator 14, and the output from the subtractor 13 is compared with a threshold.
When the output from the subtractor 13 is larger than the threshold value, it is possible to determine that the speech state is present, and when the output from the subtractor 13 is smaller than the threshold value, it is possible to determine that the speech state is not present. It becomes possible.
[0009]
[Patent Document 1]
JP-A-1-12728
[Problems to be solved by the invention]
However, in the short-time sound pressure averager 12a and the long-time sound pressure averager 12b in FIG. 3, the average time for calculating the average sound pressure of the speaker signal is different, but the average sound pressure of the speaker signal is calculated. Frequency bands are the same.
For this reason, when the utterance lasts for a long time, the steady sound pressure calculated by the long-time sound pressure averaging unit 12b is more strongly affected by the voice, so that the correct value cannot be maintained, and the accuracy of the utterance detection deteriorates. There was a problem.
[0011]
Therefore, an object of the present invention is to provide a voice detection device, a voice detection method, and a voice detection program capable of suppressing deterioration in accuracy of utterance detection due to the influence of utterance time.
[0012]
[Means for Solving the Problems]
In order to solve the above-described problem, according to the voice detection device according to claim 1, a speaker signal including a voice signal and a noise signal is input, and a level of the voice signal and a predetermined reference level are set. In the voice detection device that determines whether the speech state or the non-speech state based on the comparison result, a voice band limiting unit that band-limits the speaker signal to a frequency band in which the voice signal is distributed, First time averaging means for averaging the level of the output signal of the voice band limiting means with time, noise band limiting means for band limiting the speaker signal to a frequency band in which the noise signal is distributed, and the noise band limiting means Second time averaging means for averaging the level of the output signal of the first time averaging means for a time longer than the averaging time of the first time averaging means, and the second time averaging means based on the output signal of the first time averaging means. Subtracting means for subtracting an output signal of the stage, characterized in that it comprises a comparator for comparing the level of the level and the preset reference output signal of said subtraction means.
[0013]
This makes it possible to suppress the noise signal input to the first time averaging means and efficiently extract the audio signal, and suppress the audio signal input to the second time averaging means. , A noise signal can be efficiently extracted.
For this reason, even if the utterance lasts for a long time, it is possible to suppress the influence of the voice signal when detecting the noise level, and to improve the accuracy of the noise level calculated by the second time averaging means. Is possible, it is possible to suppress deterioration in accuracy of utterance detection due to the effect of the utterance time.
[0014]
According to the second aspect of the present invention, the level of the output signal of the second time averaging means is weighted at the subsequent stage of the second time averaging means to correct the frequency characteristic. It is characterized by further comprising a correcting means.
This makes it possible to correct the level of the output signal of the second time averaging means so that the level of the output signal of the second time averaging means matches the band of the signal input to the first time averaging means. It becomes possible.
[0015]
For this reason, even when the band of the signal input to the first time averaging means is different from the band of the signal input to the second time averaging means, the second signal is output from the first time averaging means. By subtracting the output signal of the time averaging means, noise components remaining in the output signal of the first time averaging means can be removed, and the accuracy of voice detection can be improved.
[0016]
Further, according to the voice detecting device of the third aspect, the voice band limiting unit and the noise band limiting unit may include a first high-pass filter for inputting and filtering the speaker signal, and a first high-pass filter. A second high-pass filter and a low-pass filter for filtering an output signal, wherein the speaker signal is input to the second high-pass filter via the first high-pass filter and is applied to a frequency band in which the audio signal is distributed. The limited output signal of the second high-pass filter is input to the first time averaging means, and the speaker signal is input to the low-pass filter via the first high-pass filter and the noise signal is distributed. And the output signal of the low-pass filter is input to the second time averaging means. And wherein the door.
[0017]
This makes it possible to separate the low-frequency noise band distributed in the interior of the vehicle from the audio band in which the center of the formant frequency of the audio is concentrated, and to detect the audio in a specific environment such as the interior of the automobile. Accuracy can be easily improved.
According to the voice detection method of the fourth aspect, the frequency band for extracting a voice signal from a speaker signal and the frequency band for extracting noise from the speaker signal are different from each other, so that It is characterized in that included voices are detected.
[0018]
This makes it possible to efficiently separate the voice signal and the noise signal included in the speaker signal, and reduces the influence of the noise level when detecting the voice level even when the noise signal is superimposed on the voice signal. Therefore, the accuracy of voice detection can be improved.
According to the speech detection program of the fifth aspect, the step of band-limiting the speaker signal to a first frequency band in which a speech signal is distributed, and the step of band-limiting the speaker signal in a second frequency band in which a noise signal is distributed. Band limiting the speaker signal based on a comparison result of the speaker signal band-limited to the first frequency band and the speaker signal band-limited to the second frequency band. And detecting a voice included in.
[0019]
This makes it possible to separate the speech signal and the noise signal included in the speaker signal, and to efficiently extract the speech signal and the noise signal even when the noise signal is superimposed on the speech signal.
For this reason, even when the utterance lasts for a long time, it is possible to suppress the influence of the voice signal when detecting the noise level, and it is possible to suppress the deterioration of the accuracy of the utterance detection due to the influence of the utterance time. Become.
[0020]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, a speech detection device according to an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram illustrating a schematic configuration of a voice detection device according to an embodiment of the present invention.
[0021]
In FIG. 1, a voice detection device includes an absolute value conversion circuit 3a for converting a speaker signal band-limited by the high-pass filters 1 and 2a, a low-pass filter 2b, a high-pass filter 1 and a low-pass filter 2a into an absolute value, a high-pass filter 1, An absolute value conversion circuit 3b for converting the speaker signal band-limited by the low-pass filter 2b into an absolute value, a short-time sound pressure averager 4a for calculating a short-time average sound pressure, and a long time for calculating a long-time average sound pressure A sound pressure averager 4b, a weight adjustment amplifier 5 for correcting the signal level output from the long-time sound pressure averager 4b, a subtractor 6 for subtracting the output of the weight adjustment amplifier 5 from the output of the short-time sound pressure averager 4a, and A comparator 7 for comparing the output from the subtractor 6 with a threshold value is provided.
[0022]
Here, the speaker signal includes a voice signal generated by the speaker's speech and a noise signal in a noise environment.
The high-pass filter 1 is for blocking signals in a band not transmitted by the telephone line. Since the band of the telephone line is usually 300 Hz or more, the pass band of the high-pass filter 1 is set to the band of this telephone line. Can be set corresponding to.
[0023]
The high-pass filter 2a is for extracting a signal in a band in which human formant frequencies are mainly distributed, and for cutting off low-frequency noise generated in an automobile or the like. The center of the human formant frequency is 2000 Hz. By setting the cutoff frequency of the high-pass filter 2a to around 2000 Hz to concentrate on the vicinity, it is possible to accurately extract human voice while reducing the influence of low-frequency noise.
[0024]
The low-pass filter 2b extracts a signal in a band having a weak correlation with the signal band-limited by the high-pass filter 2a. For example, the low-pass filter 2b extracts low-frequency noise generated in an automobile or the like, and also extracts a human formant frequency. Cuts off signals in a band mainly distributed.
That is, since the center of the human formant frequency is concentrated around 2000 Hz, and the noise generated in the interior of the vehicle is concentrated in a lower frequency than that, the cutoff frequency of the low-pass filter 2b is set at around 2000 Hz. Noise sound pressure can be accurately extracted while reducing the influence of the sound.
[0025]
The speaker signal from which the low-frequency component of 300 Hz or less has been cut by the high-pass filter 1 is input to the high-pass filter 2a and the low-pass filter 2b, respectively.
In the high-pass filter 2a, a signal in a band around 2000 Hz where human formant frequencies are mainly distributed is extracted, and low-frequency noise generated in the interior of an automobile or the like is cut and input to the absolute value conversion circuit 3a. You.
[0026]
When the signal extracted by the high-pass filter 2a is input to the absolute value conversion circuit 3a, the signal is converted to an absolute value and then output to the short-time sound pressure averager 4a.
Then, the short-time average sound pressure of the signal converted into the absolute value by the absolute value conversion circuit 3 a is calculated and output to the subtractor 6.
On the other hand, the low-pass filter 2b extracts low-frequency noise generated in the interior of an automobile, cuts a signal in a band around 2000 Hz where human formant frequencies are mainly distributed, and inputs the signal to the absolute value conversion circuit 3b. You.
[0027]
When the signal extracted by the low-pass filter 2b is input to the absolute value conversion circuit 3b, the signal is converted to an absolute value and then output to the long-time sound pressure averager 4b.
Then, a long-term average sound pressure of the signal converted into the absolute value by the absolute value conversion circuit 3 b is calculated and output to the weight adjustment amplifier 5.
Here, when the long-time sound pressure averager 4b is input to the weight adjustment amplifier 5, the weight adjustment amplifier 5 outputs the average sound pressure of the band output from the long-time sound pressure averager 4b to the short-time sound pressure averager. The average sound pressure output from the long-time sound pressure averager 4b is corrected so as to match the band of the signal output from the signal output unit 4a.
[0028]
The average sound pressure calculated by the short-time sound pressure averaging device 4a is input to the subtractor 6, and the average sound pressure calculated by the long-time sound pressure averaging device 4b is subtracted through the weight adjustment amplifier 5 by the subtractor 6. 6, these values are subtracted and input to the comparator 7.
Here, the weight adjustment amplifier 5 corrects the frequency characteristics of the average sound pressure output from the long-time sound pressure averaging device 4b, thereby subtracting signals band-limited to different bands, thereby obtaining a high-pass filter 2a. The remaining noise components that cannot be completely removed by the above-described method can be removed from the output signal of the sound pressure averaging device 4a in a short time, and the accuracy of voice detection can be improved.
[0029]
When the output from the subtractor 6 is input to the comparator 7, the output from the subtractor 6 is compared with a threshold.
When the output from the subtracter 6 is larger than the threshold value, it is possible to determine the utterance state, and when the output from the subtracter 6 is smaller than the threshold value, it is possible to determine the non-speech state. It becomes possible.
[0030]
FIG. 2 is a diagram illustrating a frequency distribution of voice and noise in a vehicle according to an embodiment of the present invention.
In FIG. 2, the spectral distribution A1 of the human voice is normally distributed in the range of 300 to 3500 Hz, and the center of the formant frequency of the human voice A3 exists near 2000 Hz.
[0031]
On the other hand, the peak of the in-vehicle noise spectrum distribution B1 is lower than the peak of the human voice spectrum distribution A1, and around 2000 Hz where the center of the formant frequency of the voice A3 is concentrated, the peak of the in-vehicle noise spectrum distribution B1 is lower. Corresponds to the hem.
Therefore, the filter characteristic C1 is set so that the spectral distribution B1 of the human voice is band-limited to around 2000 Hz or higher, and the filter characteristic C1 is set so that the spectral distribution B1 of the vehicle interior noise is band-limited to around 2000 Hz or lower. Set C2.
[0032]
The filter characteristic C1 is realized by combining the high-pass filters 1 and 2a, and the filter characteristic C2 is realized by combining the high-pass filter 1 and the low-pass filter 2b, whereby the average sound of the band-limited human voice spectrum distribution A2 is obtained. The pressure can be calculated by the short-time sound pressure averager 4a, and the average sound pressure of the band-limited vehicle interior noise spectrum distribution B2 can be calculated by the long-time sound pressure averager 4b.
[0033]
Therefore, it is possible to separate the voice signal and the noise signal included in the speaker signal, and to efficiently extract the voice signal and the noise signal even when the noise signal is superimposed on the voice signal.
For this reason, even when the utterance lasts for a long time, it is possible to suppress the influence of the voice signal when detecting the noise level, and it is possible to suppress the deterioration of the accuracy of the utterance detection due to the influence of the utterance time. Become.
[0034]
Here, the signal input to the short-time sound pressure averager 4a and the signal input to the long-time sound pressure averager 4b are band-limited to different bands, and the signals output from the short-time sound pressure averager 4a are When the average sound pressure calculated by the long-time sound pressure averager 4b is subtracted from the average sound pressure calculated by the short-time sound pressure averager 4a in order to remove the remaining noise, the noise band to be subtracted is different. become.
[0035]
On the other hand, as for the stationary noise in the automobile, the spectral pattern maintains a substantially constant shape regardless of the level of the noise.
For example, when the vehicle is running at a low speed, the engine sound of the vehicle is reduced, and road noise and wind noise generated between the tire and the road surface are reduced in proportion to the engine sound. Conversely, in the case of high-speed running, the engine noise of the automobile increases, and road noise and wind noise generated between the tire and the road surface increase in proportion to the engine noise.
[0036]
As a result, by multiplying the average sound pressure calculated by the long-time sound pressure averaging device 4b by the weight correction value, the noise in the band captured by the long-time sound pressure averaging device 4b is captured by the short-time sound pressure averaging device 4a. Can be converted into noise in a given band.
Therefore, the average sound pressure of the long-time sound pressure averager 4b multiplied by the weight correction value is subtracted from the average sound pressure of the short-time sound pressure averager 4a, so that the average sound pressure is input to the short-time sound pressure averager 4a. Even when the signal and the signal input to the long-time sound pressure averager 4b are band-limited to different bands, it is possible to remove noise remaining in the average sound pressure of the short-time sound pressure averager 4a. .
[0037]
That is, when the spectrum pattern maintains a substantially constant shape regardless of the level of the noise level, as shown in FIG. 2B, the integration amount of sound pressure that can be captured by the sound pressure averager 4b for a long time. The ratio between D and the sound pressure amount E that must be obtained for accurate voice detection is substantially constant.
Here, in order to input a value obtained by accurately removing noise from the voice signal included in the speaker signal to the comparator 7, the following equation (1) needs to be calculated by the subtractor 6.
[0038]
(Subtractor output) = (average sound pressure of short-time sound pressure averaging device 4a) −E (1)
Here, E is a sound pressure amount that must be obtained in order to accurately perform the sound detection in FIG.
On the other hand, regardless of the level of the noise level, when the spectrum pattern keeps a substantially constant shape, the sound pressure amount E that must be obtained in order to accurately detect the voice is captured by the long-time sound pressure averager 4b. It can be calculated by the following equation (2) using the integrated amount D of the sound pressure that can be obtained.
[0039]
E = (weight correction value) × D (2)
Therefore, by substituting equation (2) into equation (1), the subtractor output can be calculated using the following equation (3).
(Subtractor output) = (average sound pressure of short-time sound pressure averager 4a) − (weight correction value) × (average sound pressure of long-time sound pressure averager 4b) (3)
Accordingly, by inputting the average sound pressure of the short-time sound pressure averager 4a to the subtractor 6, and inputting the average sound pressure of the long-time sound pressure averager 4b to the subtractor 6 via the weight adjustment amplifier 5, It is possible to calculate a value obtained by accurately removing noise from the voice signal included in the speaker signal, and to include the noise signal in the speaker signal even when the voice signal and noise included in the speaker signal are band-limited to different bands. It is possible to input to the comparator 7 a value obtained by accurately removing noise from the audio signal to be received.
[0040]
The weight correction value can be set to an appropriate value. For example, when there is no sound or when only noise is input, the weight correction value is set so that the output from the subtractor 6 does not exceed the threshold value. Alternatively, if there is a voice when voice + noise is input, a value may be set so that the output from the subtracter 6 exceeds a threshold value.
Further, since the engine sound, road noise, and the like are unique to each vehicle type, the spectral pattern is unique to each vehicle type, but when the vehicle type changes, the weight correction set by the weight adjustment amplifier 5 is set. Can be changed, so that a spectrum pattern of any shape can be handled.
[0041]
【The invention's effect】
As described above, according to the present invention, it is possible to separate the speech signal and the noise signal included in the speaker signal by performing the band limitation of the speaker signal, and to separate the speech signal when detecting the noise level. Since the influence can be suppressed, it is possible to suppress the deterioration of the accuracy of the utterance detection due to the influence of the utterance time.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating a schematic configuration of a voice detection device according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating a frequency distribution of voice and noise in a vehicle according to an embodiment of the present invention.
FIG. 3 is a block diagram illustrating a schematic configuration of a conventional voice detection device.
[Explanation of symbols]
1, 2a High-pass filter 2b Low-pass filter 3a, 3b Absolute value conversion circuit 4a Short-time sound pressure averager 4b Long-time sound pressure averager 5 Weight adjustment amplifier 6 Subtractor 7 Comparator A1 Human voice spectrum distribution A2 After band limitation Human voice spectrum distribution A3 Human voice B1 In-vehicle noise spectrum distribution B2 In-vehicle noise spectrum distribution C1 and C2 after band limitation Filter characteristics D Sound pressure integral amount E that can be captured by long-time sound pressure averaging E Voice The amount of sound pressure that must be obtained for accurate detection

Claims (5)

A speaker signal including a voice signal and a noise signal is input, and the level of the voice signal is compared with a predetermined reference level, and it is determined based on the comparison result whether the voice state is a voice state or a non-voice state. In the voice detection device,
Voice band limiting means for band limiting the speaker signal to a frequency band in which the voice signal is distributed,
First time averaging means for averaging the level of the output signal of the audio band limiting means with time;
Noise band limiting means for band limiting the speaker signal to a frequency band in which the noise signal is distributed,
Second time averaging means for averaging the level of the output signal of the noise band limiting means for a time longer than the averaging time of the first time averaging means;
Subtraction means for subtracting the output signal of the second time averaging means from the output signal of the first time averaging means;
A sound detector comprising: a comparator for comparing a level of an output signal of the subtraction means with a preset reference level. 2. The apparatus according to claim 1, further comprising a weight correcting means for weighting an output signal level of said second time averaging means and correcting a frequency characteristic at a stage subsequent to said second time averaging means. The voice detection device according to the above. The voice band limiting means and the noise band limiting means,
A first high-pass filter for inputting and filtering the speaker signal;
A second high-pass filter and a low-pass filter for filtering an output signal of the first high-pass filter;
The speaker signal is input to the second high-pass filter via the first high-pass filter and band-limited to a frequency band in which the audio signal is distributed,
An output signal of the second high-pass filter is input to the first time averaging means,
The speaker signal is input to the low-pass filter via the first high-pass filter and band-limited to a frequency band in which the noise signal is distributed,
3. The voice detection device according to claim 1, wherein an output signal of the low-pass filter is input to the second time averaging unit. Detecting a voice included in the speaker signal by making a frequency band for extracting a voice signal from a speaker signal different from a frequency band for extracting noise from the speaker signal; . Band limiting the speaker signal to a first frequency band in which the audio signal is distributed;
Band limiting the speaker signal to a second frequency band in which a noise signal is distributed;
Detecting a voice included in the speaker signal based on a comparison result between the speaker signal band-limited to the first frequency band and the speaker signal band-limited to the second frequency band; And a step of causing a computer to execute the steps.

Images