JPH0327698A - Sound signal detection method - Google Patents

Abstract

PURPOSE:To detect a desired sound period under non-steady state noise by providing a microphone array system having a directivity control function and a sound receiver with a different S/N from the system at the same position and deciding a power difference of received sound for a time period. CONSTITUTION:A 1st sound receiver 41 outputting a signal with high S/N consists of a microphone array 51 making up of plural microphone elements and a directivity characteristic control section 52. On the other hand, a 2nd sound receiver outputs a signal with low S/N. The receivers are placed at the same location and power calculation sections 43, 44 detect the power for a time period for a short period respectively. A sound period detection section 45 detects the difference and when the value is within a prescribed range, it is decided to receive an object signal. There is a difference in the S/N in the two signals by the method and since noise and sound period are timewise matched, a desired sound period is detected under non-steady-state noise.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a sound detection method for detecting a time period in which a desired sound signal exists in a signal in which noise and a desired sound signal are mixed. [Prior art] In recent years, there has been remarkable development of speech recognition devices.
The development of noise-resistant speech recognition devices is slow. The reason for this is that it is difficult to correctly detect speech segments (determine the time segment in which speech exists on the time axis) in a noisy environment. If the ja sound section is mistakenly determined to be speech, the noise is forced to be associated with some phoneme, making it impossible to obtain correct speech recognition results.

Therefore, the development of speech segment detection technology that works well even under noisy conditions is considered to be extremely important. Figure 13 is the first
FIG. 2 is a diagram illustrating a conventional speech interval detection method. The figure is
It represents the temporal change in the short-term power of the signal, with the vertical axis representing the short-term power of the signal output from the microphone, and the horizontal axis representing time. Hereinafter, in this specification, unless otherwise specified, "power" and Lt short-time power are used. The signal includes stationary noise 11 (noise whose power is almost constant over time: for example, air conditioning noise or equipment fan noise), non-stationary noise 12 (noise whose power fluctuates greatly over time: PA
(for example, the sound of a door closing or unwanted sounds) and (desired)
Contains 13 sounds. Although it is possible to know the power of stationary noise in advance, the power of non-stationary noise is unpredictable. The first conventional method continues to monitor the power of the signal, and determines a time period in which the power becomes larger than a threshold value Thl 4 determined based on the power of stationary noise to be a speech period. Most current speech recognition devices use this method to detect speech segments.

However, with this method, the correct voice section 1 shown in Figure 3
6 can also be detected, but non-stationary noise section 1 with large power can also be detected.
5 also had a major problem in that it was incorrectly determined to be a voice section. The second conventional method to solve this problem is to use two microphones, one microphone has a large S/N ratio between the voice and ambient noise, and the other microphone has a small S/N ratio. SN on output
They are installed so that there is a difference in ratio. The specific method of installing the microphone to achieve this is shown in Figure 14 (a).
As shown in FIG. 14(b), the first microphone 1 is installed near the speaker 3, and the second microphone 2 is installed far from the speaker 3. Alternatively, as shown in FIG. A conceivable method is to install the microphone 1 in front of the speaker 3 and the second microphone 2 on the side of the speaker 3. If these installation methods are used, the audio power output from the first microphone 1 will be greater than the audio power output from the second microphone 2.
Considering that electric sounds occur far away, both microphones 1,
The noise powers at the outputs of 2 are approximately equal, so that
A difference in SN ratio occurs between the outputs of the two microphones 1 and 2. FIG. 15 is a diagram explaining the ideal operation of the second conventional method. 2 represents the temporal change in the short-term power P2 of the microphone output,
In each figure, as in FIG. 13, 11 represents stationary noise, 12 represents unsteady noise, and 13 represents voice. 2
As a result of installing the two microphones so that a difference in signal-to-noise ratio occurs, the voice power at the short-time power P2 is smaller than the voice power at the short-time power pt, while the noise power is equal in both. In the second conventional method, as shown in FIG. 15(C), the difference PD (PD=P1−
P2), and a time interval 18 in which this power difference PD is greater than a certain threshold value pth indicated by symbol 17 is determined to be a voice interval. It can be seen from FIG. 15(C) that the second conventional method does not have the problem of erroneously determining a section of large-power non-stationary noise 12 as a speech section, unlike the first conventional method.

However, in reality, this second conventional method rarely works as ideally. The reason for this is that the following three conditions must be satisfied in order to correctly detect a speech segment using the power difference between two signals. Condition 1: There is a difference in SN ratio between the two signals.

Condition 2: Both the noise section and the voice section in the two signals are temporally consistent.

Condition 3: The variation in the difference in the SN ratio due to variations in various environmental conditions is small. (Stability of difference in SN ratio) However, the second conventional method focuses only on the first condition and does not consider the second and third conditions, which causes the following problems. First, the first problem will be explained. Figure 16 is a diagram with noise source 4 added to Figure 14(a). At this time, audio is first input to the first microphone 1 and then input to the second microphone 2. On the other hand, noise is first input to the second microphone 2 and then input to the first microphone 1. Therefore, the voice section and the noise section do not match in the output signals of the two microphones. This is shown in Figure 17.
17(a) shows the short-time power P1 of the first microphone output, FIG. 17(b) shows the short-time power P2 of the second microphone output, and FIG. 17(C) shows the short-time power P1 of the first microphone output. Each represents the difference PD. Also, as in the example of FIG. 15, 11 represents stationary noise, 12 represents non-stationary noise, and 13 represents voice. The relationship between the power magnitude of the voice and the voice in FIGS. 17(a) and (b) is the same as that in FIGS. 15(a) and (b). However, in FIG. 17, the sound is transmitted at the output of the second microphone for a time τ, indicated by the symbol 3l, from the output of the first microphone.
The noise is delayed by S, and the noise is advanced by a time τN indicated by symbol 32.

In other words, both the speech interval and the noise interval are not temporally consistent. As a result, the power difference PD between the two signals is the first
As shown in FIG. 7(c), the result is different from FIG. 15<C), and the threshold value Pt. If a section of h or more is determined to be a voice section, the first problem arises in that the section indicated by symbol 33 in FIG. 17(C) is erroneously determined to be a voice section. Since the time difference τN shown by symbol 32 in this noise section varies greatly depending on the position of the noise source, it is impossible to measure the consistency using a delay device or the like. Next, the second problem is that in the actual environment, there are various factors that cause the difference in SNR between two microphone output signals to fluctuate, and it is difficult to ensure the stability of the difference in SNR between the two signals. Explain that it is difficult to do so. The first variable factor is the position of the noise source. In the above description, it is assumed that the noise source is located far away, but when the noise source is located relatively close, the location of the noise source becomes a large factor in the variation of the difference in the S/N ratio.

An example will be shown using FIG. Figure 18 <a) (b)
16, 1 and 2 are the first and second microphones, 3 is the speaker, and 4 is the same as the example shown in FIG.
is a noise source. If the noise source were in the position shown in these two figures, the power of the noise at the output of the first microphone 1 would be greater than the power of the noise at the output of the second microphone 2, similar to the power of the voice. ．． the result,
The difference in signal to noise ratio between the two microphone outputs will be small.

The second variable factor is the movement of the speaker. For example, in FIG. 18(b), when the speaker turns his or her head 45 degrees to the right, the sound is received by the two microphones with approximately the same power. As a result, there is no difference in audio power between the outputs of the two microphones 1 and 2, and the difference in SN ratio varies. The third variable factor is the influence of indoor reflected sound. 2
In many cases where two microphones 1 and 2 are installed with different signal-to-noise ratios, reflected sounds with different temporal structures and magnitudes are added to the noise and speech at each microphone, so that the signal-to-noise ratio is It fluctuates greatly. Furthermore, there are many other fluctuation factors such as electrical noise and vibration noise. Therefore, it is extremely difficult to secure a stable SN ratio difference in an environment where these SN ratio fluctuation factors exist, and it is difficult to find a microphone installation method that allows the second conventional method to work effectively. It's not easy. As described above, the second conventional method has serious problems and cannot exhibit sufficient performance in practical use.

Next, a third conventional method aimed at solving the problems of the second conventional method will be explained using FIG. 19.
In FIG. 19, 1 is the first microphone and 2 is the second microphone, as in the example described above. Also, 2
1 is a short-time power calculation unit, 22 is a voice section candidate selection unit,
23 and 24 are average power calculation units for voice section candidates;
25 is a power difference detection section, and 26 is a voice section candidate testing section. In this method, as in the second conventional method, the first microphone 1 has a large S/N ratio between voice and ambient noise, and the second microphone 2 has a small S/N ratio compared to the former microphone 1. It is installed like this. In this method, first, the short-time power calculation section 21 calculates the short-time power of the output signal of the first microphone. Next, the speech segment candidate detection unit 22 continues to monitor the short-term power of the signal, and selects a time segment whose power is greater than a threshold Th determined based on the power of stationary noise as a speech segment candidate. ．． The operation so far is the 13th
This is exactly the same as the first conventional method shown in the figure. Therefore, the noise section indicated by symbol 5 in FIG. 3 is also selected as a speech section candidate. Next, the average power calculation units 23 and 24
, the first microphone 1 in this candidate section
The average power of the output of the second microphone 2 and the average power of the output of the second microphone 2 are calculated. Next, the power difference detection section 25 calculates the difference PDL between the respective average powers. Finally, in the voice section candidate testing section 26, a predetermined threshold P
When it is larger than DLt, the candidate section is determined to be a speech section, and when it is smaller than DLt, the candidate section is rejected.

The characteristic feature of this third conventional method is that it calculates the difference in average power within a relatively long period of time selected as a speech period candidate in the output of the first microphone 1, rather than the difference in short-time power. That's true. Therefore, Fig. 17 (a
) and (b), even if the voice section and noise section are not temporally aligned in the two microphone outputs,
Furthermore, even if there is a temporal fluctuation in the S/N ratio due to the addition of reflected sounds with different temporal structures to the two signals, this has only a small effect on the difference in average power. Problems will be improved.

[Problems to be Solved by the Invention] However, since this method determines the speech section based on the average power within the candidate section, incorrect judgment results may be obtained if a noise section and a speech section exist consecutively. arise. Second
Figure 0 shows an example of such a case. FIG. 20 shows the output of the first microphone l, and the correct audio section is section 34 in the figure. In this figure, unsteady noise 1
2 and the voice 13 are temporally close to each other, a section 35 in which the short-time power exceeds the threshold Th indicated by the symbol 14 and which combines the noise section and the speech section is selected as a speech section candidate. Therefore, as a result of calculating the difference in average power, if this candidate section is determined to be the correct speech section, the second
The section indicated by symbol 36 in Figure 0 will be an erroneously determined section, and if this speech section is rejected, it will be regarded as a correct speech section or a non-speech section, so in either case, There is also the problem of incorrect judgment results. From this, it can be seen that this third conventional method does not solve the problems of the second conventional method. As described above, conventional speech interval detection methods have many of the problems mentioned above, making it difficult to detect correct speech intervals when non-stationary noise is present. Therefore, the main object of the present invention is to provide a method that can detect speech intervals in a non-stationary noise environment with higher probability than before.

Another object of the present invention is to detect a speech interval even if the noise source is located at any position other than the vicinity of the speaker (within a range of ±30 degrees when looking at the speaker from the microphone). The goal is to provide the following. [Means for Solving the Problems] In order to achieve such problems, the present invention requires the following requirements. That is, as described above, the following three conditions are necessary to correctly detect a speech section using the power difference between two signals. Condition 1: There is a difference in SN ratio between the two signals. Condition 2:
Both the noise section and the speech section of the two signals are temporally consistent. Condition 3: The variation in the above-mentioned SN ratio difference due to variations in various environmental conditions is small. (Stability of SN ratio difference) The first feature of the present invention is that in order to simultaneously satisfy the first and second conditions above, the present invention In order to operate effectively, two receivers that generate signals with different signal-to-noise ratios are installed in locations that can be considered to be virtually the same, and the difference in power between the two output signals is used to detect voice sections. It is in the point of doing.Also,
A second feature of the present invention is that, in order to satisfy the third condition, one of the two sound receivers uses a microphone array system having a directivity control function. [Operation] According to the first feature of the present invention, since both noise and speech arrive at the two receivers at the same time, both the noise section and the speech section in the output signals of the two receivers are temporally equal to each other. It is consistent. Therefore, the first problem in the second conventional method is solved. Furthermore, if the two sound receivers are installed at the same location, the temporal tR structure of the reflected sound added to each signal will also be the same, which is the second problem with the second conventional method. The influence of reflected sound on the fluctuation of the difference in the S/N ratio between the outputs of the two sound receivers is significantly reduced.

Next, according to the second feature of the present invention, the position of the noise source that affects the fluctuation of the difference in the SN ratio between the two receiver outputs, which is mentioned as the second problem in the second conventional method, and The problem of speaker movement can be improved. [Example] A configuration diagram of the present invention is shown in Fig. 1. In Figure 1, 4
Reference numeral 1 denotes a first sound receiver (microphone array system) that outputs a signal with a high signal-to-noise ratio, and is composed of a microphone array 51 composed of a plurality of microphone elements and a directional characteristic control section 52. 42 is the SN of the first receiver output
The second receiver outputs a signal with a lower signal-to-noise ratio than the signal-to-noise ratio, and these two receivers are installed at the same location. Further, 43 and 44 are short-time power calculation units, and 45 is a voice section detection unit based on the power difference between two signals.

Now, in order to explain the effects of the present invention, a unidirectional microphone is used instead of the microphone array system as the first sound receiver 4■ in the configuration shown in FIG.
Consider a method using an omnidirectional microphone as the sound receiver 42. In this way, the S/N ratio of the output of the first sound receiver, which has directivity toward the speaker, will be the same as that of the second sound receiver, which has no directivity.
The signal-to-noise ratio of the output of the receiver is greater than that of the receiver. However, this method does not always work well. This will be explained using FIG. 2. In FIG. 2, 61 is a unidirectional microphone, 62 is an omnidirectional microphone,
Each direction pattern is shown, 3 is the speaker,
63 and 64 represent the positions of miscellaneous a sources. Figure 2(a)
．． As can be seen from (b), the unidirectional microphone has high sensitivity in the front direction toward the speaker, and low sensitivity in the opposite direction. Omnidirectional microphones have the same sensitivity in all directions. Therefore, the male sound source is as shown in FIG. 2(a). At position 63 in (b), the SN ratio of the output of the unidirectional microphone is much larger than the SN ratio of the omnidirectional microphone. However, Fig. 2(a)
．． In (b), for example, when the noise source is at the position of symbol 64 (or when it moves to that position), the sensitivity of the unidirectional microphone to noise increases, so the output of the unidirectional microphone increases. The difference in the SN ratio of the output of the omnidirectional microphone becomes small. As described above, the method of using a unidirectional microphone as the first sound receiver has a problem in that the S/N ratio varies greatly depending on the position of the noise source.

The problem when using the above-mentioned unidirectional microphone; well, it can be solved by using a superdirectional sound receiver as shown in Figure 3 as the first sound receiver 41 in Figure 1. It may be thought like this. However, the directional characteristics of normal superdirectional sound receivers differ depending on the frequency. That is, in the low frequency range, it has a wide directional characteristic as shown by symbol 61 in FIG. 2(a), and in the high frequency range, it has a directional characteristic that is even sharper than that shown in FIG. 2(a). As a result, for noise in the low frequency range, there is the problem that the SN ratio fluctuates depending on the position of the noise source, as described above, but in the high frequency range, the problem is that the SN ratio fluctuates with the slightest movement of the speaker. occurs. As explained above, in order to obtain good voice section detection results, it is necessary to use a well-known directional sound receiver as the first sound receiver 4■ in the configuration of the present invention shown in FIG. It turns out that it is difficult to substitute.

Next, it will be explained that in the present invention, which uses a microphone array system with a directivity control function, fluctuations in the SN ratio can be kept small even when the position of a noise source or the speaker moves. A typical example of a microphone array system with a directivity control function is an adaptive array (micro
It is a sound receiver called a phone) array). An example of the configuration of an adaptive array is shown in FIG. In FIG. 4, 5l is a microphone array, which is composed of Mg microphone elements 561 to 56lIl. 52 is a directivity control unit, which includes filters 531 to 53M connected to each microphone output, and an adder 5 that takes the sum of the filter outputs.
5 and a filter characteristic control section 54.

The filter characteristic control unit 54 receives each microphone output signal and the output xi of the adder 55, and controls the filters 531 to 53l so as to reduce the noise component included in xi.
Control the characteristics of 4. Next, this filter characteristic control section 5
The operating principle of 4 will be explained. The output signal xi of the adder 55 is expressed as the sum of the audio component S and the noise component n as shown in the following equation.

Xl=s+n (1) At this time, if we find the filter characteristics that minimize the power n2 of the noise component without adding any conditions, the filters 531 to
53M all become filters with zero gain. As a result, the noise component n becomes zero and becomes the minimum, but the voice component S
is also not output, which is a meaningless result. Therefore,
A certain constraint condition is set for the audio component S included in the signal xi obtained as a result of the filter operation, and the characteristics of the filter that minimizes the noise component n included in xi under that condition are determined. As an example of the constraint condition, the audio component included in the microphone output signal (filter input signal) is
When expressed as
A condition that the average value of 2 is less than or equal to a predetermined threshold is known.

Now, the outputs of the M microphone elements are expressed as ul~uM, and the characteristics of the filters 53l~53lI1 are hl~hl4.
The power xl2 of the signal xl is expressed as follows.

It is expressed as M. Further, assuming that speech and noise are uncorrelated with each other, the following equation holds true.

xl2=s2+n2 (3) (2), (3
), it can be seen that the power n2 of the noise component included in xi is a quadratic function of the filter characteristic hl-hl4.
Therefore, the problem of filter control that minimizes the power n2 of the noise component under constraint conditions becomes the well-known problem of minimizing a quadratic function with constraint conditions. For various solutions and specific algorithms for various constraint conditions, please refer to the literature ("Introduction
．． oAdaptive Arrays"R.A.Mo
nzingo ef. at, John Wiley &amp;
Sons, NEW YORK, 1980) and US Pat. No. 4,536,887. In this way, reducing the noise component contained in Forms a directional characteristic with low sensitivity in the direction of the source.

Figure 5 shows one of the directional characteristics formed by the adaptive array PA6.
6 is shown. In FIG. 5, 3 is a speaker as in the previous embodiments, and 63 and 64 are noise sources. As can be seen from Figure 5, the adaptive array does not have a sharp directional characteristic, but it does achieve a directional characteristic with low sensitivity in the direction of the noise source. This low-sensitivity part of the directional characteristic is called the "blind spot".
When the microphone array is composed of M elements, the array system can account for M-1 blind spots.

An adaptive array that forms such a directional characteristic has a lower signal-to-noise ratio compared to a superdirectional receiver when a large number of noises reflected in the room arrive from directions other than the noise source. small. However, it is possible to obtain a nearly constant S/N ratio regardless of the position of the noise source, and
Because it does not have sharp directivity in the direction of
This receiver is very suitable for ensuring the stability of the ratio difference.

In addition, adaptive arrays have the characteristic of reducing temporal fluctuations in noise power. This can be seen in Figure 6(a).
This will be explained using (b). Generally, in a room, noise reflected from walls, floors, ceilings, etc. from directions other than the direction of the male sound source enters the receiver. An adaptive array cannot form a blind spot in all of these noise directions, and when a microphone array is composed of M microphone elements, a maximum of M microphone elements can be formed in the direction in which direct sound and high-energy reflected sound are incident. The signal-to-noise ratio is improved by creating a blind spot.

This effect is shown in Figure 6(a). This will be explained using (b). FIG. 6(a) shows the pulse noise signal when the sound is received by the omnidirectional microphone, and FIG. 6(b) shows the pulse noise signal when the sound is received by the adaptive array.

In FIG. 6(a), 71 is noise directly received from a noise source, and 72, 73, and 74 are noises received after being reflected one or more times from walls, floors, etc.

Compared to the energy of direct sound 71, reflected sound 72, 73,
The energy of 74 decays exponentially with time. Assuming that the number of microphone elements constituting the array is 4,
This adaptive array forms three blind spots: in the direction of the noise source and in the direction of the reflected sound at 72, 73. Therefore, the power of the reflected sound of the noise shown at 74 in the adaptive array output in FIG. 6(b) is not much different from that received by the omnidirectional microphone, but The power of sound has decreased significantly. As a result, it can be seen that the temporal fluctuations in the noise power become smaller. As mentioned earlier, the key factors for false detection of voice sections are:
This is a large temporal fluctuation in the power of the noise. In order to deal with this temporal variation, speech interval detection is performed using the power difference between the two signals, but it is impossible to completely eliminate various causes of S/N ratio variation, so false detections may occur. 1
00% cannot be avoided. Therefore, the feature of the adaptive array used in the present invention, which reduces temporal fluctuations in noise power, is very effective in reducing false detections of voice sections. Second sound receiver 4 in the configuration example of the present invention in FIG.
2, the simplest method is to use one of the microphone elements forming the microphone array 51. An example of this is shown in Figure 7 below. Further, as shown in FIG. 10, the second sound receiver inputs some of the outputs of the microphones of the microphone array 5l of the first sound receiver 42 to the signal receiver 52A and obtains the output. , it is also possible to obtain a second signal x2. Another example of a microphone array system with directional VI control is a sound receiving system as shown in US Pat. No. 791,418. In this method, signal processing is performed to preserve voice signals with a clear direction of arrival and to reduce noise arriving from the surrounding area. In order for this method to work well, it is necessary that the positions of the speaker and the noise source do not match (the directions seen from the microphone may be the same), and only the sound from the sound source located at the desired position is required. It can be considered a type of directional control in the sense that it extracts the

FIG. 7 is a diagram illustrating in more detail the first embodiment of the present invention shown in FIG. In the figure, 51 is a microphone array, 52 is a directional characteristic fil control section, 43 is a first short-time power calculation section, 44 is a second short-time power calculation section, and 45 is a voice section detection section based on a power difference. Certain things are the same as in the previous embodiments. Further, 81 is a first amplifier connected to the output side of the directional characteristic control unit 52 to receive the signal x1 and send the output to the power calculation unit 43; 82 is a first amplifier that is connected to the output side of the directional characteristic control unit 52; (using one of the microphone elements)
is connected to the power calculation unit 4 to receive the signal x2 and output the output.
83 is a power calculation unit 43;
44, a difference device that receives the outputs pi and p2; 84, a determination unit that receives the output p1 of the power calculation unit 43 and is based on the power of a short period that may be part of a voice section; 85, a difference device; 83 is a determination unit based on the power that receives the output, and 86 is the output S of the determination unit 84 based on the short-time power.
1 and the output S2 of the power-based determination section 85.

The steps to carry out this method are as follows.

First, the sound with superimposed noise is received by the microphone array 5L. The Tagata signal from the microphone array 51 is input to the directivity control section 52, which generates a first signal xi. On the other hand, the output of one microphone element constituting the microphone array 51 is assumed to be x2. At this time, as a result of the directivity control by the directivity control v4 section 52, the SN ratio at xi is larger than the SN ratio at X2. Next, using amplifiers 81 and 82, signals xi and x2 are
Correct the signal level so that the power of the audio contained in is equal. Although this operation is not essential, performing this operation will simplify the explanation later. next,
In the short-time power calculation units 43 and 44, X
Short-term powers P1 and P2 of 1 and x2 are calculated and output. This short-time power pt and P2 are logarithmic (d
B) or expressed as an exact value.

Next, the power P1 of the signal with a high SN ratio is input to the power-based determination section 84. In the determination unit 84 based on this power, if the value of P1 is larger than a predetermined leap value Th, an output St is output to indicate the possibility that the corresponding short period is part of a vocal section. 1” is output, otherwise it outputs “0”.

Next, in the differentiator 83, the difference PD (
PD=P2-PL> is calculated, and this difference PD is input to the determination unit 85 based on the power difference.

The determination unit 85 based on this power difference outputs "1" as the output S2 when the value of PD is smaller than a predetermined threshold pth, and otherwise outputs ``0''''. .

Finally, the output S1 of the determination section 84 based on the power and the output S2 of the determination section 85 based on the power difference are input to the voice section determination section 86. In the speech section determination unit 86, when the values of S1 and S2 are both "1", the short time section that has become a candidate is determined to be part of a correct speech section, and in other cases, it is determined that it is a noise section. Outputs the determined result.

Next, the operation of the voice section detection section 45 based on the power difference will be explained using FIGS. 8(a), (b), and (c). FIG. 8(a) shows the power PL at the output of the first receiver.
8(b) represents the temporal change in the power P2 at the output of the second receiver, and FIG. 8(C
) represents the difference PD between P2 and P1 (PD=.P2-P1). In each figure, the vertical axis represents the short-time power of the signal, and the horizontal axis represents time. Further, 11 represents stationary noise, 121 and 122 represent non-stationary noise, and 13 represents voice in the same manner as described in the above example.

The power of the voice contained in pt and P2 is adjusted to be equal, so if the power of the stationary noise in P2 is somewhat smaller than the power of voice, the power of the voice contained in pt and P2 is adjusted to be equal. In Figures (a) and (b), the powers of the voice sections are approximately equal. on the other hand,
Since the output of the second receiver has a smaller S/N ratio than the output of the first receiver, the power of the noise in FIG. 8(b) is smaller than the power of the noise in FIG. 8(a). It is shown that the difference is increased by an amount corresponding to the difference in the S/N ratio. As a result, P2 and pt shown in FIG. 8(C)
The value of the power difference PI) is zero in the speech section, and takes a non-zero value in the non-speech section. However, in a real environment, various SN
Since there are factors that cause variation in the ratio difference, even in the present invention, which uses a microphone array system with a directivity control function to reduce the variation factors, the PD value does not take such an ideal value. is not limited. For example, if the speaker moves within a larger range than expected, the PD value will be greater than zero even in the voice section, and noise coming from the same direction as the voice (for example, the speaker's tongue clicking, For the sound of a speaker turning a paper, etc., even if the sound has relatively low power, the PD value will be zero in that noise section. Taking this point into consideration, Therefore, in the present invention, first, as shown in FIG. 8(a), the power-based determination unit 84 operates to determine that the pheasant time interval smaller than the threshold Th is a non-speech interval.As a result, for example, , even if the noise indicated by symbol 122 is noise coming from the same direction as the voice and the PD is small in that noise section, this noise section will not be mistakenly detected as a speech section, and the effectiveness will be reduced. It can be seen that high speech interval detection is achieved.

The voice section determination unit 86 shown in FIG. 7 outputs the output s from the power-based determination unit 84 as shown in FIG.
1 and the output S2 from the determination unit 85 based on the power difference are both “1”, this testing unit determines that the short time interval is a voice interval in addition to the voice interval candidate test unit 86a that determines the short time interval as a voice interval. An interval verification unit 86b may be provided that determines the time interval to be a voice interval only when the time interval continues beyond the predicted value of the minimum duration interval of voice.

In order to confirm the effectiveness of the present invention, the following experiment was conducted. <Experimental conditions> The experiment was conducted in a room with a reverberation time of 0.4 seconds f,:.

As for the noise, interference sound (radio news) was generated from the Subika. The desired audio is word audio (city name)
We collected 100 words uttered under different distracting voices. The positions of the speaker and the noise source were set at 45 degrees apart from the receiver. The sound receiver 1 is an A M N O. which is one of the adaptive arrays. R sound receiving device (Reference: Y. Kaneda and J. Ohga ”A
dapt. ive Microphone-array
SystCmfor Noise Reduction
n, fEEE'rrans. on Acous
t. ．． ,Speech,Signal Processes
sing, vol-ASSP-34, PP. 1391-
1400, Dec. 1986) was used. Al4N
The OR receiver a device is realized by combining a microphone array composed of multiple microphone elements and a digital filter, and has a lower lO than a single microphone element.
It is possible to receive sound with a high SN ratio of ~16 dB. Also,
As the sound receiver 2, two microphone elements, which are the constituent elements of the aforementioned microphone array, were used. Short-term power calculations were performed every 10 ms with a window length of 30 ms. The threshold Th in the power-based determination unit 84 is determined by capturing each utterance for a certain length (1 second), finding the difference Pl4M between the maximum short-time power and the minimum short-time power, and determining Th
= PMM X O. 5.

Further, the PD threshold PLh was set to 8 dB.

Note that, as the correct answer for the speech section, the section obtained by applying the first conventional method (method using only determination based on power) to speech in the absence of noise was used.

(Experimental Results) Under the above conditions, a word section detection experiment was conducted with the SN ratio of the voice at the sound receiving point set to -5 dB at the output of the sound receiver 2.

FIG. 9 shows an example of the experimental results. Figure 9(a) shows the correct answer for the speech power and speech section when there is no noise. FIG. 9(b) shows the power P2 of the output of the second receiver when interfering sound is added.

Fig. 9(C) shows the power PL of the output of the first sound receiver (AMNOR sound receiving device) and the selected sound section candidates when the interfering sound is added. The part indicates the incorrectly detected voice section.
When comparing Figures (b) and (C), the temporal fluctuation of the noise power indicated by the Δ symbol in Figure (b) is smaller in Figure (C), which is the output of the adaptive array. I understand that.

In other words, the sharp peak of the temporal change in power has become flat.

FIG. 9(d) shows the results determined to be word intervals as a result of applying the method of the present invention with arrows. In addition, Fig. 9 (
The non-speech section within 200 S between the speech sections detected in c) and (d) was considered to be part of the word section. The hatched portions are sections that were erroneously detected (a speech section was determined to be a noise section). From this figure, it can be confirmed that the method of the present invention operates almost satisfactorily.

In order to quantitatively evaluate the experimental results, a case where the error at the start and end of a word section could be detected within 50 Ils was regarded as a correct answer, and the correct answer rate was calculated. When the first conventional method, which is the most commonly used method in current speech recognition devices, is applied to the output of AMNOH, which has a high S/N ratio, the accuracy rate is 43? ≦. On the other hand, in the method of the present invention, a detection result of 96% was obtained, and the average detection error at the beginning and end was about 20ffis. These results confirmed the effectiveness of the present speech interval detection method.

Furthermore, if a unidirectional microphone is used as the first sound receiver, as shown in FIG. 2(a), for example,
If a noise source exists within a range of 90 degrees from the microphone in the direction of the speaker, with respect to the straight line connecting the speaker and the microphone, the correct answer rate for the word section is approximately 10%; It was confirmed that the invention is a highly accurate acoustic signal detection method. In addition, in the present invention, the above-mentioned test result of ±96% is obtained except for the range of ±30° with respect to the straight line connecting the speaker and the microphone.

For applications where a slight performance deterioration can be tolerated, a sound receiver composed of a so-called superdirectional sound receiver and a selection filter can also be applied as the first sound receiver of the present invention. FIG. 12 shows an example of its configuration. In FIG. 12, 51 is a microphone array, 91 is an adder for realizing superdirectivity, 9
2 is a processing filter. As mentioned above, when using a superdirectional sound receiver, the SN is low in the low frequency range and high frequency range.
Since the fluctuation in the ratio becomes large, this processing filter has high sensitivity in the range where the speaker is expected to move, and improves this problem by extracting only the band with low sensitivity outside of that range. It is. The problem with this method is S
Since the frequency band with small fluctuations in the N ratio does not necessarily match the band with high voice energy, the S/N ratio of the output of the first sound receiver decreases, and the number of incorrect selections in voice section candidates increases. ．． On the other hand, the advantage of this method is that the system configuration is simple.

The present invention does not utilize any characteristics specific to audio signals. However, in order to perform voice section detection, it is very effective to use a determination method that utilizes the properties of voice signals in combination with the present invention.

In fact, the first conventional method is not used alone, but is usually used in combination with a determination method that utilizes the properties of the audio signal. For example, a method is known in which a predicted value Tc of the minimum duration of a voice signal is used to determine that candidates for a voice section shorter than Tc are noise. Combining this judgment method to remove the influence of pulsed noise is
This is a very effective method for detecting voice sections. In addition, many other determination methods are known, such as a method that uses the periodicity of the audio signal to determine that sections where the signal is aperiodic are non-speech. These conventional methods input the section determined to be a speech section by the present invention and re-determine the section, or finalize the speech section by majority vote as a result of multiple determinations including the present invention. Thus, the present invention can be combined with many conventionally known speech interval detection methods, so that
Depending on the purpose of use, it is also possible to achieve significant improvements in detection performance. Now, the first field of application of the present invention is application to speech recognition devices, as explained above. The second field of application is the acoustic echo canceller. Acoustic echo canceller is a technology used, for example, in public address telephone systems, to prevent the sound from the receiving speaker from reaching the transmitting microphone and being received, resulting in problems such as howling. The principle of an acoustic echo canceller is to estimate the acoustic transfer characteristics from the receiving speaker to the transmitting microphone, and then subtract the sound component from the receiving speaker from the signal received by the transmitting microphone based on the estimation results. be. Since the transfer characteristics from the receiving speaker to the transmitting microphone change over time, it is necessary to continuously estimate the transmission characteristics, but the condition is that the speaker is not speaking at the time of estimation (otherwise, a large estimation error) is necessary. However, it is not always possible to determine whether the speaker is speaking or not, and this is one of the current issues with this technology.

To solve this problem, the present invention is applied by considering the voice of the sender as the target voice and the voice from the receiver speaker as the unnecessary voice. If it is assumed that the person is making a sound and the estimation operation of the transfer characteristic is stopped, it becomes possible to realize a high-performance acoustic echo canceller that solves the above problems.

The third application field is application to voice storage technology. For example, when trying to digitize a large amount of continuous vocalizations and record them on a magnetic disk, etc., information compression technology using audio encoding is also important, but it is also possible to detect non-speech sections and truncate them, or It is also a very important technology to record information with a particularly low amount of information. The present invention is applicable to detecting non-speech intervals in such techniques. Furthermore, since the method of the present invention does not utilize the inherent properties of audio signals, it is possible to select any sound other than audio (for example, music, mechanical sounds, impact sounds, etc.) as the sound to be detected. ．． As a result, the method of the present invention can be applied to various types of monitoring devices, measuring devices, etc. [Effects of the Invention] As explained above, the method of the present invention allows the sound to be received by the first sound receiver (microphone array system with directional control function) and the second sound receiver installed at the same location. Since the presence of a desired signal is determined by using the short-term power difference between the audible signals, it is possible to detect a desired voice section in a non-stationary noise environment, which was impossible with conventional methods of this type. This makes it possible.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining an embodiment of the acoustic signal detection method according to the present invention, and FIG. 2 is a diagram for explaining problems when using a unique directional microphone and an omnidirectional microphone. Fig. 3 is a diagram for explaining the problems when using a super-directional receiver, Fig. 4 is a block diagram showing a specific example of the first receiver in Fig. 1, and Fig. 5 is Figure 6 is a diagram showing the directivity characteristics of the adaptive array. Figure 6 is a waveform diagram showing the received signal waveform of a pulsed voice when an omnidirectional microphone and adaptive array are used. Figure 7 is shown in Figure 1. FIG. 8 is a block diagram illustrating the embodiment shown in FIG. 7 in more detail, and FIG.
Figure 1 shows the experimental results that confirmed the effectiveness of the present invention.
Figures O through 12 are block diagrams showing other embodiments of the present invention, Figure 13 is a graph showing the first example of the conventional voice-containing interval detection method, and Figure 14 is a graph showing the conventional voice interval detection method. A diagram showing an example of microphone installation for explaining the second example, FIG. 15 is a graph for explaining the ideal operation of the second conventional method,
FIG. 16 is a graph showing the positional relationship between the microphone and the noise source, FIG. 17 is a graph for explaining the problem of the second conventional method, and FIG. 18 is a graph showing the positional relationship between the microphone and the noise source. FIG. 19 is a block diagram showing a third example of the conventional speech interval detection method, and FIG.
This is a graph for explaining the problem of the example. 41.41...Speech receiver, 43.44...Short-time power calculation unit, 45...Voice section detection unit, 5■...
... Microphone array, 52 ... Directivity control section, 8
4.85...determination section, 86...voice section determination section.

Claims (9)

[Claims] (1) Using first and second receivers that are installed at approximately the same location and transmit signals with different power ratios of target signal and noise (SN ratio), these receivers are used to receive these sounds in a certain time interval. If the power difference or ratio of the signals transmitted from the receiver is within a predetermined range, it is determined that the target signal has been received in this time interval, and the first receiver , an acoustic signal detection method characterized in that it is an adaptive microphone array whose directional characteristics can be controlled according to the noise position. (2) The acoustic signal detection method according to claim 1, wherein the first and second sound receivers use sound receivers having different directivity characteristics. (3) The acoustic signal according to claim 1, wherein the first sound receiver is composed of a microphone array composed of a plurality of microphone elements and a directional characteristic control circuit arranged at a subsequent stage. Detection method. (4) In claim 1, said 2 in a certain time interval
The difference or ratio of the power of the two signals is within a predetermined range, and the power of the signal output from the receiver with a high SN ratio in a certain time interval is within the predetermined range. If so, it is determined that the target signal has been received during this time interval. (5) The acoustic signal detection method according to claim 1, wherein the second sound receiver is also constituted by a microphone array. (6) In claim 1, if the time interval in which it is determined that the target signal has been received continues beyond the predicted minimum duration of audio, it is determined that the target signal has been received in this time interval. An acoustic signal detection method characterized by: (7) The acoustic signal detection method according to claim 3, wherein the second sound receiver uses one microphone element that is a component of a microphone array that constitutes the first sound receiver. . (8) In claim 6, the first sound receiver includes a microphone array composed of a plurality of microphone elements and a directional characteristic control circuit disposed at a subsequent stage, and the first sound receiver The acoustic signal detection device is characterized in that the device shares some of the microphone elements constituting the microphone array constituting the first sound receiver, and further has means for synthesizing the outputs of these several microphone elements. Method. (9) Using first and second receivers that are installed at approximately the same location and transmit signals with different power ratios of target signal and noise (SN ratio), these receivers are used to receive these sounds in a certain time interval. If the power difference or ratio of the signals transmitted from the receiver is within a predetermined range, it is determined that the target signal has been received during this time interval, and the first receiver , a directional microphone array in which a plurality of microphones are arranged, a synthesizer that receives the output of each microphone and synthesizes superdirectivity, and the output of this synthesizer is passed through a receiver's predetermined band components and is processed by a band selection filter. An acoustic signal detection method comprising: