JP2003223175A - Sound block detector - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a sound block detector by which a sound block can be surely detected even in the sound signal of a low S/N. <P>SOLUTION: The sound signal collected by a microphone 11 and amplified by a line amplifier 12 is converted to a digital value by an A/D 101 and stored in a storage part 102. After a noise is removed, the digitized sound signal improves the S/N by taking a self-correlation in a short time and when the period of the S/N greater than a threshold is continued for a prescribed period, the sound block is decided. Further, in order to surely detect the head and end of the sound block, the preceding part and the following part of the sound block are forcibly opened for a predetermined period. Besides, in order to prevent the threshold from becoming too great by storing noises, the prescribed multiple of a moving average value for the predetermined period of a non-sound block is successively updated as a threshold. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice section detecting device, and more particularly to a voice section detecting device capable of surely detecting a voice section even for a voice signal having a low signal-to-noise ratio.

[0002]

2. Description of the Related Art In voice recognition, it is necessary to accurately extract a voice section to be voice-recognized from a voice signal containing noise captured through a microphone.

Conventionally, a voice section detection method for determining a voice section when a state in which the voice level is equal to or higher than a predetermined threshold value continues for a predetermined time or longer is generally applied. It is difficult to secure sufficient accuracy when trying to recognize many kinds of words from a voice signal.

In order to solve the above problems, the present applicant has already proposed a voice section detecting device for detecting a voice section based on a voice pitch signal.

[0005]

However, the voice pitch-based voice section detection device can reliably detect the voice section even for a word including a consonant, or for a word in which the Sa or Ha sound is continuous. Although it is possible, when the voice level of the speaker is low, for example, when the speaker is a female, the voice pitch can be extracted at the beginning or end of the voice section because a sufficient signal-to-noise ratio cannot be secured. This is impossible and it is difficult to detect the voice section.

The present application has been made in view of the above problems, and an object of the present invention is to provide a voice section detection device capable of surely detecting a voice section even for a voice signal having a low signal-to-noise ratio. And

[0007]

According to a first aspect of the present invention, there is provided a voice section detecting device comprising a pre-processing means for removing noise contained in a voice signal, and a signal pair of the voice signal from which noise has been removed by the pre-processing means. Signal-to-noise ratio improving means for improving the noise ratio,
A voice section extraction signal generating means for generating a voice section extraction signal based on the voice signal whose signal to noise ratio has been improved by the signal to noise ratio improving means is provided.

In the present invention, after removing the noise,
The signal-to-noise ratio is improved, and the voice segment extraction signal is generated based on the voice signal with the improved signal-to-noise ratio.

In the voice section detecting apparatus according to the second aspect of the present invention, the signal-to-noise ratio improving means calculates the short-time autocorrelation value calculating means for calculating the short-time autocorrelation value of the voice signal from which noise has been removed by the preprocessing means. Is.

In the present invention, the short-time autocorrelation value of the voice signal is calculated in order to improve the signal-to-noise ratio of the voice signal.

According to a third aspect of the present invention, a voice section detecting device is provided when a state where the short time autocorrelation value calculated by the short time autocorrelation value calculating means is equal to or more than a predetermined threshold value continues for a predetermined time. The speech segment extraction signal is opened.

According to the present invention, the voice segment extraction signal is opened when the state in which the short-time autocorrelation value is equal to or greater than the predetermined threshold value continues for the predetermined time period.

In the voice section detection device according to the fourth aspect of the present invention, the voice section extraction signal generating means uses a threshold value which is a multiplication value of the average level of the voice signal when the voice section extraction signal is closed and a predetermined scaling factor. Includes threshold setting means for setting.

In the present invention, the threshold value is sequentially updated as a predetermined multiple of the average value of the noise level included in the voice signal.

In the voice section detecting device according to the fifth aspect of the present invention, the level of the short time autocorrelation value calculated by the short time autocorrelation value calculating means by the voice section extraction signal generating means is equal to or higher than a predetermined threshold value. Extraction signal opening means for opening the extraction signal when the state continues for a predetermined time, and when the extraction signal is opened by the extraction signal opening means, the extraction signal is opened retrospectively for a predetermined period. The extraction signal retroactive opening means for outputting the extracted signal as a voice section extraction signal is provided.

According to the present invention, the extracted signal is opened when the state in which the short-time autocorrelation value is equal to or greater than the predetermined threshold value continues for a predetermined time, and the extracted signal is opened retrospectively for a predetermined period. Is output as a voice segment extraction signal.

In the voice section detection device according to the sixth aspect of the present invention, the voice section extraction signal generating means preliminarily sets a state in which the short time correlation value calculated by the short time autocorrelation value calculating means is equal to or more than a predetermined threshold value. Extraction signal opening means for opening the extraction signal when the extraction signal is opened for a predetermined time, and extraction signal is predetermined even after the extraction signal is closed when the extraction signal is opened by the extraction signal opening means An extraction signal open maintaining unit is provided for outputting a signal maintained for a period of time as a voice section extraction signal.

In the present invention, the extraction signal is opened and the extraction signal is closed when the state in which the level of the short-time autocorrelation value is equal to or higher than a predetermined threshold value continues for a predetermined time. The signal maintained in the open state for the subsequent predetermined period is output as the voice section extraction signal.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a functional block diagram of a voice section detecting device according to the present invention.
An audio signal converted into an electric signal by the microphone 11 and amplified by the line amplifier 12 is taken in.

The voice section detection device 10 comprises an analog / digital conversion section (A / D) 101, a storage section 102, a voice signal processing section 103, a voice section extraction signal generation section 104, and a voice section extraction section 105.

That is, the audio signal is sampled in the A / D 101 every predetermined sampling time T seconds and stored in the storage unit 102.

The voice section extraction signal generation unit 104 generates a voice section extraction signal based on the output of the voice processing unit 103,
The voice section extraction unit 105 extracts a voice section from the digitized voice signal stored in the storage unit 102 based on the voice section extraction signal.

In this embodiment, the A / D 10
1, the storage unit 102, the voice signal processing unit 103, the voice section extraction signal generation unit 104, the voice section extraction unit 105 is configured by using a personal computer (PC), the voice signal processing unit 103, the gate signal generation unit 104, The voice section output unit 105 is configured as software, and the program is stored in the PC.
Installed in to function as a voice section detection device.

FIG. 2 is a flow chart of a main routine of a program which is recorded in a storage medium such as a CD-ROM and is installed in the PC.
In step A, the audio signal to be processed is sampled by the A / D 101 for a predetermined sampling time and stored in the storage unit 102. Although the sampling time can be set as appropriate, in the present embodiment, the sampling time T = 0.08333 milliseconds (sampling frequency = 1
2 kHz).

In step 21, an initial value setting routine for initializing parameters used for voice processing, a voice processing routine for improving the signal-to-noise ratio of the voice signal in step 22, and a voice-processing routine after improving the signal-to-noise ratio in step 23. A voice section extraction signal generation routine for generating a voice section extraction signal based on the voice signal is executed.

Finally, in step 24, a voice segment extraction routine for extracting a voice segment from the voice signal stored in the storage unit 102 is executed based on the voice segment extraction signal, and this routine is ended.

FIG. 3 is a flow chart of the initial value setting routine executed in step 21.
In, the parameter of the high-pass filter used in the audio signal processing routine is set to the initial value based on [Equation 2].

[0028]

[Equation 2]

Next, in step 211, the parameters of the low pass filter are set based on [Equation 3].

[0030]

[Equation 3]

Then, in step 212, the parameters used in the short-time autocorrelation routine are set in step 21.
At 3, the parameters used in the root mean square routine are initialized.

Next, in step 214, the parameters used in the smoothing routine are initialized by [Equation 4], [Equation 5] and [Equation 6].

[0033]

[Equation 4]

[0034]

[Equation 5]

[0035]

[Equation 6]

Further, in step 215, the parameters used in the voice section extraction signal generation routine are initialized, and this routine is ended.

FIG. 4 is a flow chart of the audio signal processing routine executed in step 22 of the main routine. In step 220, the parameter n representing the sampling point is set to the initial value "0".

Using the parameters of the high-pass filter determined in step 210 of the parameter setting routine in step 221, the high-pass filter is applied to the audio signal X _i (n) stored in the storage unit 102 based on [Equation 7]. Execute the filter routine to execute the high-pass filtered signal X
Output _h (n).

[0039]

[Equation 7]

This processing is for removing the noise of the air conditioner radiated into the passenger compartment, and the cutoff frequency f _ch of the high pass filter is selected to be 300 hertz, for example.

Next, in step 222, the low-pass filter routine is executed on the high-pass filter output signal X _h (n) based on [Equation 8] using the parameters of the low-pass filter determined in step 211 of the parameter setting routine. And low-pass filtered signal X _L (n)
Is output.

[0042]

[Equation 8]

This processing is for removing high frequency noise that is suddenly generated, and the cutoff frequency f _cL of the low pass filter is selected to be 3000 hertz, for example.

Then, in step 223, in order to improve the signal / noise ratio, a short-time autocorrelation routine is executed on the low-pass filter output signal X _L (n) to calculate the short-time autocorrelation signal X _c (n). To do.

Next, in step 224, the root mean square value X _p (n) of the short-time autocorrelation signal X _c (n) is calculated, and in step 2
At 25, the root mean square value X _p (n) is smoothed by a low pass filter to calculate a smoothed output X _s (n).

Further, in step 226, the smoothed output X
Perform a gate routine on _s (n) to obtain a gate signal G
Calculate (n).

Then, in step 227, it is determined whether or not the calculation of the gate signal G has been completed for the N audio signals X _i , and if a negative determination is made, the parameter n is incremented in step 228 and the steps from step 221 onward are performed. Repeat the process.

On the contrary, when the affirmative determination is made in step 227, that is, when the audio signal processing is completed for the N audio signals X _i , this routine is ended.

The details of the processing of steps 223 to 226 will be described below.

FIG. 5 shows step 223 of the voice processing routine.
2 is a flowchart of a short-time autocorrelation routine executed by the low-pass filter processing audio signal XL (n) separated by a predetermined number of samples M and a low-pass filtered audio signal separated by a predetermined number of samples M from the signal XL (n). By calculating a predetermined number J of correlation values between the signals _XL (n-M) based on [Equation 9], the signal level in the voice section is relatively set with respect to the noise level in the non-voice section. Enlarge.

[0051]

[Equation 9]

First, at step 2230, it is determined whether the current sampling point n is equal to or more than the sum of the separation number M and the correlation number J. The distance M and the correlation number J are set in step 212 of the parameter setting routine.

When an affirmative decision is made in step 2230,
That is, when the current sample point n is more than the sum of the number of samples M and the number of correlations J and the autocorrelation can be calculated, the process proceeds to step 2231 and the parameter j representing the number of integrations and the integration value S are set to "0". Initialization to step 223
2 again and S the sum of the integrated value and S of X _L (n-j) and _{X L (n-j-M} ) at.

In step 2233, it is determined whether the parameter j is greater than or equal to the correlation number J, and when a negative determination is made, that is, when the parameter j is less than the correlation number J, the parameter j is incremented in step 2234 and step 2232 is performed. Repeat the process.

When an affirmative decision is made in step 2233,
That is, when the parameter j becomes equal to or larger than the correlation number J, the integrated value S is divided by the correlation number J in step 2235 to calculate the short-time autocorrelation signal X _c (n), and this routine is finished.

When the negative determination is made in step 2230, that is, when the current sample point n is separated and is less than the sum of the number of samples M and the number of correlations J, the autocorrelation cannot be calculated. Time-divided correlation signal X
This routine is ended by setting _c (n) to "0".

Here, the separation number M and the correlation number J must be experimentally determined so that the voice section can be accurately detected regardless of the speaker, but the separation number M is 3
It is desirable to set it so as to correspond to milliseconds (for example, 36 when the sampling time is 0.08333 milliseconds), and it is desirable that the number of correlations J is 5.

FIG. 6 is an explanatory view of the effect of the short-time autocorrelation processing, in which (a) shows the signal X after the low-pass filter processing.
_L (n), (B) shows a voice signal waveform obtained by shifting this waveform by a time interval (= 3 milliseconds), and (C) shows a waveform of the short-time autocorrelation signal X _c (n). From this figure, it can be understood that the signal-to-noise ratio is improved by performing the autocorrelation processing for a short time.

FIG. 7 shows step 224 of the voice processing routine.
2 is a flow chart of a root mean square routine executed in step 1, wherein root mean square processing is performed on the short time autocorrelation signal X _c (n) in order to eliminate the influence in the amplitude direction of the short time autocorrelation processing signal X _c. Run.

First, in step 2240, it is determined whether or not the current sampling number n is equal to or less than a predetermined number N _p (for example, 200). If a positive determination is made, in step 2241, the root mean square processed signal X _p (N) is set to "0" and this routine ends. This is to remove noise at the beginning of the short-time autocorrelation signal X _c (n).

When a negative decision is made in step 2240,
That is, when the removal of the top portion is completed, step 2242
It is determined whether or not the parameter k has reached a predetermined value K (for example, 32) which has been set in advance, and when a negative determination is made, the addition value of the squared value of S and X _c (n) is changed to S again in step 2243. deep.

Next, at step 2244, the root mean square signal X _p
(N) is set to the hold signal X _po and the parameter k
Is incremented and this routine ends.

When an affirmative decision is made in step 2242,
That is, when the parameter k reaches the predetermined value K, the value obtained by dividing the integrated value S by J is squared in step 2245 to calculate the root mean square signal X _p (n), and the holding output X _po is calculated as the root mean square signal X _p. Set to (n).

Then, in step 2246, the parameter S
And k are reset and this routine is finished.

When the root mean square processing is completed, in order to remove the high frequency component (particularly the impulse component) contained in the root mean square signal Xp in step 225 of the audio signal processing routine, the fifth-order low-pass represented by [Equation 10] is used. Smooth using IIR filter.

[0066]

[Equation 10]

FIG. 8 is an explanatory view of the effect of the smoothing process.
Short-time autocorrelation signal X shown in (d) _cTwo for (n)
The root mean square signal X shown in FIG.
_pIt was found that (n) contains a considerable high frequency component.
It Therefore, when further smoothing processing is performed, the smoothing shown in (F)
Signal X_s(N) is smooth and the threshold value can be easily determined.
I understand.

FIG. 9 shows step 226 of the voice processing routine.
It is a flowchart of the gate routine executed in step 2, in which a gate opening / closing routine is executed in step 2260 and a threshold value setting routine is executed in step 2261.

FIG. 10 shows step 226 of the gate routine.
In the flowchart of the gate opening / closing routine executed at 0, the threshold value TL is set to a predetermined value TR of the noise level ZL (n-1) one sample before at step 60a (for example, 1.8).
It is set to double, and it is determined in step 60b whether the smoothed signal X _s (n) is less than or equal to the threshold value TL. When n = 0, the initial value of the noise level one sample before is initialized to "0" in step 215 of the initial value setting routine.

When an affirmative decision is made in step 60b, that is, when the smoothed signal X _s (n) is below the threshold value TL, the gate G signal (n) at the current sample point is set to "1" (open) in step 60c. Set and exit this routine.

On the contrary, when a negative determination is made in step 60b, that is, when the smoothed signal X _s (n) is larger than the threshold value TL, in step 60d the gate signal G of the current sampling point is obtained.
(N) is set to "0" (closed) and this routine ends.

FIG. 11 shows step 226 of the gate routine.
2 is a flowchart of a threshold value setting routine executed in step 2. Since the voice level has individual differences, it is difficult to detect a voice section that does not depend on the speaker when the threshold value is a fixed value. Therefore, the threshold value is automatically updated. To do.

That is, the average value of the root-mean-square signal X _{p in} the non-voice section in which no voice is present is set as the noise level, and the predetermined magnification of this noise level is set as the threshold value. However, if the number of samples for which the average value is calculated is unlimited, there is a risk that the threshold value will be kept high due to the influence of high-level noise that occurred a long time ago. The number of root mean square signals X _p to be set is limited to a predetermined number M (for example, 1200).

FIG. 12 is an explanatory diagram of a voice section and a non-voice section.
And the square mean signal X of (h) _pIs greater than the threshold
The section (section b) is the speech section, and the root mean square signal X_pThe threshold
Sections smaller than the value (sections a and c) are non-voice sections
It The gate signal G (n) does not open in the section b.
ing.

That is, in step 61a, the gate signal G
If (n) is "0", and a positive determination is made, that is, if there is no voice, step 61b.
Then, it is determined whether or not the parameter m is equal to or less than the predetermined number M of targets for noise level calculation.

When a positive determination is made in step 61b, that is, when the parameter m is less than the predetermined value M, step 61c
Then, the root mean square signal X _p (n) is added to the noise integrated value ZT to update the noise integrated value ZT.

Next, at step 61d, the root mean square signal X
_p (n) is held in the holding root mean square signal X _po (n), and the parameter m is incremented in step 61e.

Then, in step 61f, the noise integrated value ZT
Is divided by a predetermined number M (for example, 1200) to set the noise level ZL (n), the noise level holding value ZLB is updated with the current noise level ZL (n) in step 61g, and this routine is finished.

The processing in steps 61f and 61g is prepared in case the gate G signal (n + 1) of the next sampling number shifts to "1".

On the contrary, when a negative determination is made in step 61b, that is, when the parameter m is equal to or larger than the predetermined value M, the holding root mean square signal X is calculated from the noise integrated value ZT in step 61h.
Subtract _po (0). This is 1 for the average value calculation target
Since it is limited to 200, this is a process for removing the oldest value X _po (0) of the smoothing signal for holding X _so (1), and setting it as 1199 integrated values before updating the noise integrated value ZT.

Next, at step 61i, a shift process for advancing the holding mean square signal X _po by one is executed. The details of the shift process will be described later.

Next, at step 61j, the current root mean square signal X _p (n) is added to the noise integrated value ZT to set the number of additions to M, and the noise integrated value ZT is updated. At step 61k, the noise integrated value ZT is predetermined. Noise level ZL divided by the number M
Set to (n).

Then, in step 61m, the noise level holding value is updated with the current noise level ZL (n), and this routine ends.

If a negative determination is made in step 61a, that is, if it is a voice section, the current noise level ZL (n) is the noise level holding value ZLB which is the noise level calculated in the immediately preceding non-voice section in step 61n. To end this routine.

FIG. 13 shows step 61 of the threshold setting routine.
i is a flowchart of the shift routine executed in i.
Then, in step 61i0, the parameter m_pInitial to "0"
And the holding root mean square signal X in step 61i1._po(M
_p+1) to X_po(M_p) As holding mean square signal X_poTo
Shift forward. Parameter m in step 61i2 _p
Is less than a predetermined value "M-1", and a positive determination is made.
Parameter 61, the parameter m_pThe ink
Then, the process of step 61i1 is repeated.

On the contrary, when a negative determination is made in step 61i2, that is, when the parameter m _p reaches the predetermined value "M-1", the current mean square value signal X is calculated in step 61i4.
Let _P (n) be the M-1th holding root mean square signal X _Po (M
-1) is kept and this routine is finished.

As described above, the processing of the voice signal processing routine in step 22 of the main routine is completed, and then the voice section extraction signal generating routine is executed in step 23 of the main routine.

FIG. 14 is a flow chart of the voice section extraction signal generation routine executed in step 23 of the main routine. In step 230, the basic extraction signal generation routine for generating the extraction signal which is the basis of the voice section extraction is
In step 231, a front addition routine for retroactively opening the basic extraction signal is executed, and in step 232, a rear addition routine for maintaining the opening state for a predetermined time after the basic extraction signal is closed is executed.

FIG. 15 is a flow chart of the basic extraction signal generation routine executed in step 230 of the voice section extraction signal generation routine, in the case where the open state continues for a predetermined period when the gate is opened in the gate opening / closing routine. It is a routine for determining that the basic voice section.

First, in step 2300, a parameter n (parameter indicating a sampling point) used in this routine, F (a flag indicating whether or not the open transition process has been executed), and i (a sampling for continuing the open state). Reset the parameter for counting points).

In step 2301, it is determined whether the gate signal G (n) set in the gate opening / closing routine is "1 (open)", and if a positive determination is made, step 230
At 2, the parameter i is incremented.

In step 2303, it is determined whether or not the parameter i is equal to or larger than a predetermined number I (for example, 480), and I ensures that the gate signal G (n) maintains the "1 (open)" state. It is a number corresponding to the time during which it can be determined that it is in the voice section, and if it is 40 milliseconds and the sampling time is 0.08333 milliseconds, I will be 480.

When an affirmative decision is made in step 2303,
That is, when the open state of the gate signal G (n) continues for the predetermined number I, the open processing routine is executed in step 2304, which will be described in detail later.

When the open processing routine is completed, step 2
At 305, it is determined whether the parameter n is less than the total number N of sampling points. When the determination is affirmative, that is, when the processing has not been completed for all the sample points, the parameter n is incremented in step 2306 and the processing of steps 2301 to 2304 is repeated.

On the contrary, when a negative determination is made in step 2305, that is, when the processing is completed for all sampling points, this routine is ended.

When a negative decision is made in step 2301,
That is, when the gate signal G (n) is “0 (closed)”,
The extraction signal E (n) is set to zero, the parameters F and i are reset, and the process proceeds to step 2306.

When a negative determination is made in step 2303, that is, when the number i of open states of the gate signal G (n) is equal to or less than the predetermined number II, the extraction signal E (n) is set to zero and the parameter is set to zero. Step 23 after resetting F
Proceed to 06.

FIG. 16 is a flowchart of the open processing routine executed in step 2304 of the basic extraction signal generation routine, and it is determined in step 4a whether the flag F is "1".

When an affirmative decision is made in step 4a, that is, when the open transition processing has already ended, the current extraction signal E (n) is set to "1" in step 4b and this routine ends.

Conversely, when a negative decision is made in step 4a,
That is, when the open transition processing is not completed, the gate signal G
(N) is "1", but the open transition processing 4c to 4g is performed to retroactively set "1" to the extracted signal E set to "0" because it has not continued II pieces.

That is, in step 4c, the parameter j representing the retrospective sampling number is reset, and in step 4d, the extracted signal E (n-j) traced back j times from the present is set to "1".

Next, in step 4e, it is determined whether or not the parameter j is equal to or greater than the predetermined number II, and when a negative determination is made, that is, when the retroactive process is not completed, the parameter j is incremented in step 4f and the process proceeds to step 4d. Return.

Conversely, when an affirmative decision is made in step 4e,
That is, when the retroactive process is completed for the predetermined number of samplings, the flag F is set to "1" and this routine is completed.

FIG. 17 is a flowchart of the forward addition routine executed in step 231 of the voice section extraction signal generation routine. Considering that the voice level is generally low at the start point of voice, the voice section is considered. In order to reliably detect the beginning of the, the extracted signal E is retroactively extended forward for a predetermined period.

That is, in step 2310, the parameter n (parameter indicating the sampling point) and FB (flag indicating whether or not the forward addition processing is completed) used in this routine, and i (the number of sample points for continuing the open state) are set. Parameter to be counted) is reset.

Next, at step 2311, the extracted signal E (n)
Is "1 (open)", and if the result is affirmative, the front addition processing routine is executed in step 2312 and the process proceeds to step 2314. Conversely, when a negative determination is made in step 2311, that is, the extraction signal E (n) is “0.
If it is (closed) ", the flag FB is set to" 0 "in step 2313 and the process proceeds to step 2314.

In step 2314, it is determined whether or not the parameter n is less than the total number N of sampling points, and when a positive determination is made, that is, when the processing has not been completed for all sampling points, the parameter n is incremented in step 2315. Returning to step 2311. On the contrary, when a negative determination is made in step 2314, that is, when the processing is completed for all the sample points, this routine is ended.

FIG. 18 shows step 23 of the front addition routine.
12 is a flowchart of the forward addition processing routine executed in step 12, wherein the current sample point n is less than the number of samples NB corresponding to a period in which the basic extraction signal should be extended forward (for example, 50 milliseconds) in step 12a? To judge.

When a positive determination is made in step 12a, that is, when the extraction signal E (0) at the beginning to the extraction signal E (n-1) immediately before the present is set to "1", the process proceeds to step 12b. move on. In step 12b, it is determined whether or not the front addition processing has been completed, that is, whether the flag FB is "1". If a negative determination is made, the parameter j indicating the number of retroactive samples is set to n in step 12c.

Then, in step 12d, the extracted signal E (j
-1) is set to "1" and it is determined in step 12e whether the parameter j is "1".

When a negative decision is made in step 12e, the parameter j is decremented in step 12f and the process of step 12d is repeated. On the contrary, if the affirmative decision is made in step 12e, it is considered that the front addition process has been finished, and the flag FB is set to "1" in step 12g, and this routine is finished.

If a negative determination is made in step 12a, that is, if the extraction signal E (n-NB) to the extraction signal E (n-1) immediately before the current one is set to "1", step 12h is executed. move on. In step 12h, it is determined whether the front addition processing has been completed, that is, whether the flag FB is "1". If a negative determination is made, the parameter j representing the number of retroactive samples is set to NB in step 12i.

Then, in step 12j, the extracted signal E (j
-1) is set to "1", and it is determined in step 12k whether the parameter j is "1".

When a negative decision is made in step 12k, the parameter j is decremented in step 12m and the process of step 12j is repeated. On the other hand, if the affirmative decision is made in step 12k, it is considered that the front addition process is completed, and the flag FB is set to "1" in step 12g, and this routine is completed.

Incidentally, step 12b or step 12h
If the affirmative determination is made in step S21, that is, if the forward addition process has already been completed, the current value of the extraction signal E (n) is maintained at "1", and the flag FB is set to "1" at step 12g. Exit the routine.

FIG. 19 is a flowchart of the rearward addition routine executed in step 232 of the voice section extraction signal generation routine. Considering that the voice level is generally low at the end point of voice, the voice section is considered. In order to reliably detect the end of the, the extracted signal E is extended backward for a predetermined period.

In step 2320, the parameter n (parameter representing the sampling point) used in this routine is set to "0", and in step 2321 it is determined whether the parameter n is "0".

When a negative decision is made in step 2321,
That is, when processing sampling points other than the first sampling point, it is determined in step 2322 whether the previous extraction signal E (n-1) is larger than the current extraction signal E (n).

When an affirmative decision is made in step 2322,
That is, when the extraction signal E transits from "1 (open)" to "0 (closed)", it is determined in step 2323 whether the sum of the parameter n and the predetermined number NB is less than the total number N of samples. Here, NB is the number of samples corresponding to the period for extending the extracted signal backward. If this period is 100 milliseconds, NB = 1200 when the sampling time is 0.08333 milliseconds.

When a negative decision is made in step 2323,
That is, when the number of samples to be extended backward exceeds the total number of samples, the open maintenance routine is executed in step 2324 to set "1 (open)" from the extraction signals E (n) to E (N). This routine ends.

On the other hand, when an affirmative decision is made in step 2323, that is, when the number of signals to be extended backward does not exceed the total number of samples, the extraction signals E (n) to E (n + NA) are set to "1".
In order to set "open", a halfway open maintenance routine is executed in step 2325, and the flow proceeds to step 2326.

In step 2326, it is determined whether or not the parameter n is less than the total number N of sampling points, and when a positive determination is made, that is, when the processing for all sampling points is not completed, the parameter n is incremented in step 2327. Then, the processing from step 2321 is repeated.

When the affirmative judgment is made in step 2321, that is, when the head data is processed, step 232
In step 8, the extraction signal E (n) is set to "0", and this routine ends. When a negative determination is made in step 2322, that is, the extraction signal E changes from "1 (open)" to "0 (closed)".
When the transition is to a time other than the current extraction signal E
No processing is performed to maintain the value of (n), and the process proceeds directly to step 2326.

FIG. 20 shows step 23 of the rear addition routine.
24 is a flowchart of the open maintaining routine executed in 24, in which the parameter j is reset in step 24a, and the extraction signal E (n + j) is set to "1 (open)" in step 24b.

Next, in step 24c, it is determined whether or not n + j is less than the total number N of samples, and when the determination is affirmative, that is, "1 (open)" has not been set until the final extraction signal E (N). In this case, the parameter j is incremented in step 24d and the process returns to step 24b.

On the contrary, when a negative determination is made in step 24c, that is, when the setting to "1 (open)" is completed up to the final extraction signal E (N), this routine is ended.

FIG. 21 shows step 23 of the rear addition routine.
It is a flowchart of the halfway open maintenance routine executed in 24, wherein the parameter j is reset in step 25a, and the extraction signal E (n + j) is set to "1" in step 25b.
(Open) ”.

Next, in step 25c, it is determined whether or not j is less than a predetermined number NA, and when a positive determination is made, that is, when the NA extraction signals E have not been set to "1 (open)", the step is performed. The parameter j is incremented in 25d and the process returns to step 25b.

On the contrary, when a negative determination is made in step 25c, that is, when the NA extraction signals EA have been set to "1 (open)", the parameter n is set to NA in step 25e.
Increase and end this routine.

With the above, the voice section extraction signal generation routine of the main routine is completed, and the voice section extraction signal E is generated.

FIG. 22 is an explanatory diagram of the effect of the forward addition and backward addition processing. When the opening / closing is determined by comparing the root mean square signal X _p with the threshold value, the gate signal G is Since the opening and closing are frequently repeated, it is not possible to accurately extract the voice section.

On the other hand, when the gate signal G is subjected to the front addition and the rear addition as described above,
As shown in (e), the voice segment extraction signal is open between the 37446 sampling point and the 57591 sampling point where the voice exists.

Note that (a) in (a) is removed from the voice section extraction signal because the open duration of the gate signal G is 40 milliseconds or less.

Finally, in step 24 of the main routine, the extraction signal E is "1 (open)" by synchronizing and integrating the audio signal Xi (n) and the extraction signal E (n) stored in the storage section. It is possible to extract the audio signal Xi in a certain section.

23 and 24 are explanatory views of a voice signal processing process in the voice section detecting apparatus according to the present invention.
( _L ) is the waveform of the original signal X _i (n) of "ice cream" pronounced by a woman in the car, (O) is the waveform of the signal X _L (n) after high-pass filtering, and (W) is the low-pass filter. The waveform of the processed signal X _h (n), ( _f ) is the waveform of the short-time autocorrelation signal X _c (n).

Further, (Yo) is the root mean square signal X _p (n)
, (T) shows the waveform of the smoothed signal X _s (n), (L) shows the waveform of the gate signal G (n), and (S) shows the waveform of the voice section extraction signal E (n).

The extracted voice section can be input to a subsequent device, for example, a voice recognition device, and used to improve the voice recognition rate.

[0138]

According to the voice section detecting device of the first aspect of the present invention, the voice section extraction signal is generated based on the voice signal having the improved signal-to-noise ratio, so that the signal-to-noise ratio is low. It is possible to reliably detect the voice section.

According to the voice section detecting device of the second invention, it is possible to improve the signal-to-noise ratio of the voice signal by the short time autocorrelation value of the voice signal.

According to the speech segment detection apparatus of the third invention, the speech segment extraction signal is opened when the state in which the level of the short-time autocorrelation value is equal to or higher than the predetermined threshold value continues for the predetermined period. By so doing, it becomes possible to reliably detect the voice section even in a situation where the signal-to-noise ratio is poor.

According to the voice section detection device of the fourth aspect of the present invention, it becomes possible to successively update the threshold value.

According to the voice section detecting device of the fifth aspect of the present invention, the beginning of the voice section can be reliably detected by making the extraction signal retroactively open for a predetermined period as the voice section extraction signal. It will be possible.

According to the voice section detecting device of the sixth aspect of the present invention, the end of the voice section is surely secured by using the voice section extraction signal which is kept open for a predetermined period after the extraction signal is closed. It becomes possible to detect.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice section detection device according to the present invention.

FIG. 2 is a flowchart of a main routine.

FIG. 3 is a flowchart of an initial value setting routine.

FIG. 4 is a flowchart of an audio signal processing routine.

FIG. 5 is a flowchart of a short time autocorrelation routine.

FIG. 6 is an explanatory diagram of an effect of short-time autocorrelation processing.

FIG. 7 is a flowchart of a root mean square routine.

FIG. 8 is an explanatory diagram of an effect of smoothing processing.

FIG. 9 is a flowchart of a gate routine.

FIG. 10 is a flowchart of a gate opening / closing routine.

FIG. 11 is a flowchart of a threshold setting routine.

FIG. 12 is an explanatory diagram of a voice section and a non-voice section.

FIG. 13 is a flowchart of a shift routine.

FIG. 14 is a flowchart of a voice segment extraction signal generation routine.

FIG. 15 is a flowchart of a basic extraction signal generation routine.

FIG. 16 is a flowchart of an open processing routine.

FIG. 17 is a flowchart of a front addition routine.

FIG. 18 is a flowchart of a front addition processing routine.

FIG. 19 is a flowchart of a rear addition routine.

FIG. 20 is a flowchart of an open maintenance routine.

FIG. 21 is a flowchart of a halfway open maintenance routine.

FIG. 22 is an explanatory diagram of effects of front addition and rear addition.

FIG. 23 is an explanatory diagram (1/2) of a voice signal processing process in the voice section detection device according to the present invention.

FIG. 24 is an explanatory diagram (2/2) of a voice signal processing process in the voice section detection device according to the present invention.

[Explanation of symbols]

10 ... Voice section detection device 101 ... A / D 102 ... storage unit 103 ... Audio signal processing unit 104 ... Voice section extraction signal generation unit 105 ... Voice section extraction unit 11 ... Microphone 12 ... Line amplifier

─────────────────────────────────────────────────── ─── Continuation of front page (51) Int.Cl. ⁷ Identification code FI theme code (reference) G10L 3/00 513Z (72) Inventor Osamu Iwata 1-2-2 Goshodori, Hyogo-ku, Hyogo Prefecture Fujitsu Ten Co., Ltd. (72) Inventor Masataka Nakamura 1-1, Miyake, Saiki-ku, Hiroshima City, Hiroshima Prefectural school Tsuru Gakuen (72) Inventor Kazuya Terao 1-1, Miyake, Saiki-ku, Hiroshima City, Hiroshima Prefecture Incorporated Tsuru Gakuen F-term (reference) 5D015 AA01 CC05 DD02 DD03 DD04 DD05 EE05

Claims (10)

[Claims] 1. Pre-processing means for removing noise contained in a voice signal; signal-to-noise ratio improving means for improving the signal-to-noise ratio of the voice signal from which noise has been removed by the pre-processing means; A voice section detection device comprising voice section extraction signal generating means for generating a voice section extraction signal based on a voice signal having a signal-to-noise ratio improved by a noise-to-noise ratio improving means. 2. The short-time autocorrelation value calculation means for calculating the short-time autocorrelation value of the speech signal from which noise has been removed by the preprocessing means, based on [Equation 1]. The voice section detection device according to claim 1. [Equation 1] 3. The voice activity detection apparatus according to claim 1, wherein the preprocessing means comprises a high-pass filter for blocking low-frequency noise in a voice signal and a low-pass filter for blocking high-frequency noise in a voice signal. 4. A state in which the voice section extraction signal generating means has a state in which the level of the voice signal whose signal-to-noise ratio improving means has improved the signal-to-noise ratio is equal to or higher than a predetermined threshold value for a predetermined time. The speech segment detection device according to claim 1, wherein the speech segment extraction signal is opened when the speech segment extraction signal is continued. 5. The state in which the level of the short-time correlation value calculated by the short-time autocorrelation value calculation means is equal to or higher than a predetermined threshold value is maintained by the voice section extraction signal generation means for a predetermined time period. The voice section detection device according to claim 2, wherein the voice section extraction signal is opened at times. 6. The voice section extraction signal generating means includes a threshold value setting means for setting a threshold value to be a multiplication value of an average level of the voice signal and a predetermined scaling factor when the voice section extraction signal is closed. Item 4. The voice section detection device according to Item 4 or 5. 7. The root mean square value calculating means for calculating the root mean square value of the short time autocorrelation values calculated by the short time autocorrelation value calculating means, and the mean square value calculation. Smoothing means for smoothing the mean square value of the short-time autocorrelation values calculated by the means, and the square of the short-time autocorrelation values smoothed by the smoothing means when the speech segment extraction signal is closed. The voice section detection device according to claim 5, further comprising a threshold value setting unit that sets a threshold value to a multiplication value of an average value and a predetermined scaling factor. 8. The voice section extraction signal generation means, when the short time autocorrelation value calculated by the short time autocorrelation value calculation means is equal to or more than a predetermined threshold value for a predetermined time. When the extraction signal is opened by the extraction signal opening means for opening the extraction signal, the extraction signal is opened by the extraction signal opening means retrospectively for a predetermined period as a voice section extraction signal. The voice section detection device according to claim 2, further comprising a retrospective opening means for outputting the extracted signal. 9. The speech section extraction signal generating means, when the short-time autocorrelation value calculated by the short-time autocorrelation value calculating means is equal to or more than a predetermined threshold value for a predetermined time. And an extraction signal opening means for opening the extraction signal, and when the extraction signal is opened by the extraction signal opening means, the extraction signal is kept open for a predetermined period even after the extraction signal is closed. 3. The voice section detection device according to claim 2, further comprising extraction signal open maintaining means for outputting a signal as a voice section extraction signal. 10. The voice section extraction signal generating means, when the short time autocorrelation value calculated by the short time autocorrelation value calculating means is equal to or more than a predetermined threshold value for a predetermined time. Extraction signal opening means to open the extraction signal to, when the extraction signal is opened by the extraction signal opening means, the extraction signal retrospective opening means to open the extraction signal retrospectively for a predetermined period, When the extraction signal is retrospectively opened by the retrospective opening means, the extraction signal that is kept open for a predetermined period even after the retrospectively opened extraction signal is closed is output as a voice section extraction signal. The voice section detecting device according to claim 2, further comprising a signal open maintaining unit.