JP2913105B2 - Sound signal detection method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sound detection method for detecting a time section in which a desired sound signal exists, for a signal in which noise and a desired sound signal are mixed.

[Related Art] In recent years, there has been remarkable development of a speech recognition device, but development of a noise-resistant speech recognition device has been delayed.
The reason is that it is difficult to correctly detect a speech section (determining a time section where speech exists on the time axis) in a noise environment. If the noise section is erroneously determined to be speech, the noise is forcibly associated with some phoneme, so that it is impossible to obtain a correct speech recognition result. Therefore, the development of a speech section detection technique that operates well even under noise is considered to be very important.

FIG. 13 is a diagram for explaining a first conventional voice section detection method. This figure shows the temporal change of the short-term power of the signal, the vertical axis represents the short-term power of the signal output from the microphone, and the horizontal axis represents the time. Hereinafter, in this specification, “power” indicates short-time power unless otherwise specified. Signals include stationary noise 11 (noise whose power is almost constant over time: for example, air conditioning noise and equipment fan noise), and non-stationary noise 12 (noise whose power greatly fluctuates over time: for example, the sound of a door closing or (Unwanted voice) and (desired) voice 13 are included. The power of stationary noise can be known in advance, but the power of non-stationary noise is unpredictable. In the first conventional method, monitoring of the power of a signal is continued, and a time section in which the power becomes larger than a threshold Th14 determined based on the power of the stationary noise is determined as a voice section. Most current speech recognizers use this method to detect speech segments.

However, in this method, the correct speech section 16 shown in FIG. 13 is used.
Can be detected, but there is a serious problem that the non-stationary noise section 15 having a large power is erroneously determined to be a speech section. A second conventional method for solving this problem uses two microphones, one of which has a large S / N ratio between voice and ambient noise, and the other has a small S / N ratio, that is, two microphones. It is installed so that a difference in SN ratio occurs in the output. As a concrete installation method of the microphone for realizing this, as shown in FIG. 14 (a), the first microphone 1 is located near the speaker 3 and the second microphone 2 is located far from the speaker 3. 14 (b), or the first microphone 1 may be installed in front of the speaker 3, and the second microphone 2 may be installed in the side of the speaker 3, as shown in FIG. 14 (b). If these installation methods are performed, the first microphone 1
Assuming that the sound power outputted from the second microphone 2 is larger than the sound power outputted from the second microphone 2, while the noise is uttered at a distant place, the noise powers at the outputs of the two microphones 1 and 2 are almost equal. A difference in the SN ratio occurs between the outputs of the microphones 1 and 2.

FIG. 15 is a diagram for explaining the ideal operation of the second conventional method.
FIG. 15 (a) shows the short-time power of the output of the first microphone
FIG. 15 (b) shows the temporal change of P1 and FIG. 15 (b) shows the temporal change of the short-time power P2 of the second microphone output. In each figure, as in FIG. Non-stationary noise, 13 represents voice. Two microphones
As a result of installation so that a difference in SN ratio occurs, short-time power P2
Is lower than the power of the voice at the short-time power P1, while the power of the noise is equal in both. In the second conventional method, as shown in FIG. 15 (c), short-time powers P1 and P2 of two signals are used.
Is calculated (PD = P1−P2), and the power difference PD
Is to determine the time section 18 in which the threshold value Pth indicated by the symbol 17 has become larger than the threshold value Pth as the voice section. From FIG. 15 (c), it can be seen that the second conventional method does not cause a problem that the section of the non-stationary noise 12 having a large power is erroneously determined to be a voice section unlike the first conventional method.

However, in practice, this second conventional method rarely operates in such an ideal manner. The reason is that the following three conditions need to be satisfied in order to correctly detect a voice section using the power difference between two signals.

Condition 1: The two signals have a difference in SN ratio.

Condition 2: the noise section and the voice section in the two signals are
Both must be time aligned.

Condition 3: The change in the difference in the S / N ratio due to the change in various environmental conditions is small. (Stability of S / N ratio difference) However, the second conventional method pays attention only to the first condition and does not consider the second and third conditions, so that the following problems occur.

First, the first problem will be described. Fig. 16 shows the fourth
The noise source 4 is added to FIG. At this time,
Sound is first input to the first microphone 1 and then to the second microphone 1.
Is input to the microphone 2. On the other hand, noise is first input to the second microphone 2 and then to the first microphone 1. Therefore, in the output signals of the two microphones, the voice section and the noise section do not match.

This is shown in FIG. FIG. 17 (a) shows the short-time power P1 of the output of the first microphone, and FIG.
The short-term power P2 of the microphone output of Fig. 17 (c)
Represents the short-term power difference PD. Also, reference numeral 11 denotes stationary noise, 12 denotes non-stationary noise, and 13 denotes voice, as in the example of FIG.

The relationship between the magnitudes of voice and noise power in FIGS. 17 (a) and (b) is the same as that in FIGS. 15 (a) and (b). However, in FIG. 17, the sound is delayed at the output of the second microphone from the output of the first microphone by the time τS indicated by the symbol 31, and the noise is advanced by the time τN indicated by the signal 32. It has become something. That is, both the voice section and the noise section are not temporally matched. As a result, the difference PD between the powers of the two signals is different from that in FIG. 15 (c) as shown in FIG. 17 (c). Causes the first problem that the section indicated by the symbol 33 in FIG. 17 (c) is erroneously determined to be a voice section. The time difference indicated by the symbol 32 of this noise section τN is
Since it greatly changes depending on the position of the noise source, it is impossible to measure the consistency using a delay device or the like.

Next, as a second problem, in an actual environment, 2
Explain that there are various factors that change the S / N ratio difference between two microphone output signals, and it is difficult to ensure the stability of the S / N ratio difference between two signals.

The first of the fluctuation factors is the position of the noise source. In the above description, the noise source is assumed to be far away, but when the noise source is relatively close, the position of the noise source causes a large variation in the S / N ratio difference. An example is shown with reference to FIG. 18 (a) and 18 (b), similarly to the example of FIG. 16 described above, reference numerals 1 and 2 denote first and second microphones, respectively, reference numeral 3 denotes a speaker, and reference numeral 4 denotes a noise source. The noise source is this 2
In the positions shown in the figures, the noise power at the output of the first microphone 1 becomes larger than the noise power at the output of the second microphone 2 as in the case of the audio power. As a result, the SN between the two microphone outputs
The difference in ratio is small.

The second variation factor is the movement of the speaker. For example, when the speaker turns his head to the right by 45 ° in FIG. 18B, the sound is received by the two microphones with almost the same power. As a result, there is no longer any difference in audio power between the outputs of the two microphones 1 and 2, and the difference in SN ratio fluctuates.

As a third variation factor, there is an influence of room reflected sound.
In many cases where the two microphones 1, 2 are installed with different SNRs, reflected sounds of different temporal structure and loudness are added to the noise and speech at each microphone, so that the SNR is It fluctuates greatly over time.

Furthermore, there are many other fluctuation factors such as electric noise and vibration noise. Therefore, it is extremely difficult to secure a stable difference between the S / N ratios in an environment in which these fluctuation factors of the S / N ratio exist, and it is difficult to find a microphone installation method that can effectively operate the second conventional method. It's not easy.

Thus, the second conventional method has a serious problem,
Practically, it cannot exhibit sufficient performance.

Next, a third conventional method aiming at solving the problems of the second conventional method will be described with reference to FIG.
In FIG. 19, as in the example described above, reference numeral 1 denotes a first microphone, and 2 denotes a second microphone. Further, 21 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 unit, and 26 is a voice section candidate test unit.

In this method, as in the second conventional method, the first microphone 1 has a large S / N ratio between the voice and the ambient noise, and the second microphone 2 has a higher SNR than the former microphone 1.
It is installed so that the SN ratio becomes small. In this method, first, the short-time power of the output signal of the first microphone is calculated by the short-time power calculator 21. next,
The voice section candidate detection unit 22 continues to monitor the short-time power of the signal, and selects a time section in which the power becomes larger than a threshold Th determined based on the power of the stationary noise as a voice section candidate. The operation so far is exactly the same as that of the first conventional method shown in FIG. Therefore, the noise section indicated by the symbol 15 in FIG. 13 is also selected as a speech section candidate. Next, the average power calculators 23 and 24 calculate the average power of the output of the first microphone 1 and the average power of the output of the second microphone 2 in this candidate section. Next, in the power difference detection section 25, a difference PDL between the respective average powers is obtained. Finally, the voice section candidate test section
In 26, if the threshold is larger than a predetermined threshold value PDLt, the candidate section is determined as a voice section, and if the threshold is smaller, the candidate section is rejected.

What is characteristic in the third conventional method is not the difference in short-term power but the difference in average power in a relatively long time section selected as a speech section candidate in the output of the first microphone 1. That is. Therefore, as shown in FIGS. 17 (a) and (b), in the two microphone outputs, even if the speech section and the noise section are not temporally matched, the two signals have different temporal structures. Even if there is a temporal change in the SN ratio due to the addition of the reflected sound, the influence on the difference in the average power is small, and the problem of the second conventional method is improved.

[Problems to be Solved by the Invention] However, in this method, since the voice section is determined based on the average power in the candidate section, an erroneous determination result is obtained when the noise section and the voice section exist continuously. Occurs. No.
Fig. 20 shows an example of such a case. FIG. 20 shows the output of the first microphone 1, and the correct voice section is the section 34 in the figure. In this figure, since the nonstationary noise 12 and the voice 13 are close in time, the short-time power is
A section 35 that exceeds the threshold Th indicated by 14 and combines the noise section and the voice section into one is selected as a voice section candidate. Therefore, as a result of calculating the difference in the average power, if this candidate section is determined to be a correct voice section, the section indicated by the symbol 36 in FIG. 20 becomes an erroneously determined section, and If the section is rejected, it is determined that the voice section is correct, and the voice section is regarded as a non-voice section. In any case, an incorrect determination result occurs.

This indicates that the third conventional method is not a method for solving the problems of the second conventional method.

As described above, the conventional voice section detection method has the above-described various problems, and it has been difficult to detect a correct voice section when non-stationary noise exists.

Therefore, a main object of the present invention is to provide a method capable of detecting a speech section in a non-stationary noise environment with a higher probability than before.

Another object of the present invention is to provide a method for detecting a voice section even when a noise source is present at any position except for the vicinity of a generator (a range of ± 30 degrees when a speaker is viewed from a microphone). Is to provide.

[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 voice section using the power difference between two signals.

Condition 1: The two signals have a difference in SN ratio.

Condition 2: the noise section and the voice section in the two signals are
Both must be time aligned.

Condition 3: The change in the difference in the S / N ratio due to the change in various environmental conditions is small.

(Stability of S / N ratio difference) The first feature of the present invention is that, in order to simultaneously satisfy the above first and second conditions, the same place (not the same place in a strict sense, the present invention) In order to operate effectively, two sound receivers that generate signals with different S / N ratios are installed at places where they can be regarded as substantially the same), and the detection of a voice section is performed using the power difference between the two output signals. The point is to do. Also,
According to a second feature of the present invention, 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 voice reach the two receivers at the same time, both the noise section and the voice section in the two receiver output signals are temporally different. Be consistent. Therefore, the first problem in the second conventional method is solved. Further, if the two sound receivers are installed at the same position, the temporal structure of the reflected sound given to each signal becomes the same, so that the second problem in the second conventional method is as follows. The effect of the reflected sound on the fluctuation of the difference between the S / N ratios in the two receiver outputs described above is greatly reduced.

Next, according to a second aspect of the present invention, a noise source position affecting the variation of the S / N ratio difference between the two receiver outputs described as the second problem in the second conventional method, and The problem of speaker movement can be improved.

Embodiment FIG. 1 shows a configuration diagram of the present invention. In FIG.
Reference numeral 41 denotes a first sound receiver (microphone array system) that outputs a signal having a high SN ratio, and includes a microphone array 51 including a plurality of microphone elements and a directivity control unit 52. 42 is compared to the SN ratio of the first receiver output
This is a second sound receiver that outputs a signal with a low SN ratio, and these two sound receivers are installed in the same place. Reference numerals 43 and 44 denote short-time power calculators, and reference numeral 45 denotes a voice section detector based on a power difference between two signals.

Now, in order to explain the effect of the present invention, a unidirectional microphone is used instead of the microphone array system as the first sound receiver 41 in the configuration of FIG.
Consider a method using an omnidirectional microphone as the sound receiver 42 of FIG. By doing so, the SN ratio of the output of the first sound receiver having directivity directed to the speaker becomes larger than the SN ratio of the output of the second sound receiver having no directivity.

However, this method does not always work well. This will be described with reference to FIG. In FIG. 2, 61 indicates a directional pattern of a unidirectional microphone, 62 indicates a directional pattern of an omnidirectional microphone, 3 indicates a speaker, and 63 and 64 indicate positions of noise sources. As can be seen from FIGS. 2 (a) and 2 (b), the unidirectional microphone has high sensitivity in the front direction toward the speaker's law and low sensitivity in the opposite direction. Omnidirectional microphones have the same sensitivity in all directions. Therefore, if the noise source is located at the position of the symbol 63 in FIGS. 2A and 2B, the S / N ratio of the output of the unidirectional microphone is much larger than that of the non-directional microphone. However, in FIGS. 2A and 2B, when the noise source is at, for example, the position of the symbol 64 (or moves to that position), the sensitivity of the unidirectional microphone to noise increases. ,
The difference between the S / N ratio of the output of the unidirectional microphone and the output of the omnidirectional microphone becomes small. in this way,
The method using the unidirectional microphone as the first sound receiver has a problem that the SN ratio greatly varies depending on the position of the noise source.

The problem in the case of using the above unidirectional microphone can be solved by using a sound receiver having super directivity as shown in FIG. 3 as the first sound receiver 41 in FIG. Might be considered. However, the directional characteristics of a normal super directional sound receiver differ depending on the frequency. That is, in the low frequency range, the directional characteristic has spread like the symbol 61 in FIG. 2A, and in the high frequency range, it has a sharper directional characteristic than that shown in FIG. 2A. As a result, for low-frequency noise, as described above, depending on the position of the noise source, SN
The problem that the ratio fluctuates is a problem that the SN ratio fluctuates with a slight movement of the speaker in the high frequency range.

As described above, in order to obtain a good voice section detection result, a well-known directional sound receiver is used as the first sound receiver 41 in the configuration of the present invention shown in FIG. It turns out to be difficult.

Next, in the present invention using the microphone array system having the directivity control function, it will be described that the fluctuation of the SN ratio can be kept small even with respect to the position of the noise source and the movement of the speaker.

A representative example of a microphone array system having a directivity control function is an adaptive array (Adaptive (microphone) arra
This is a sound receiver called y). FIG. 4 shows a configuration example of the adaptive array. In FIG. 4, reference numeral 51 denotes a microphone array, which is composed of M microphone elements 56l to 56M.

52 is a directivity control unit, filters 53l to 53M connected to each microphone output, and an adder that calculates the sum of the filter outputs
55 and a filter characteristic control unit 54.

Each microphone output signal and the output xl of the adder 55 are input to the filter characteristic control unit 54, and the characteristics of the filters 531 to 53M are controlled so as to reduce noise components included in xl.

Next, the operation principle of the filter characteristic control unit 54 will be described. The output signal xl of the adder 55 is expressed by the following equation as the sum of the voice component, s, and the noise component n.

xl = s + n (1) At this time, if a filter characteristic for minimizing the power n2 of the noise component is obtained without any condition, the filters 53l to 53M
Are all zero gain filters. As a result, the noise component n becomes zero and minimizes, but the result is meaningless unless the voice component s is also output. Therefore, a certain constraint condition is set for the audio component s included in the signal xl obtained as a result of the filter operation, and under the condition, the characteristic of the filter that minimizes the noise component n included in xl is obtained. Examples of the constraint condition when representing the sound component contained in the microphone output signal (filter input signal) and s0, and constraint that s = s0, | s-so | average value of ² is preset threshold value or less Are known.

Now, if the outputs of the M microphone elements are represented by u1 to uM and the characteristics of the filters 531 to 53M are represented by h1 to hM, the signal x1
Power x1 ² of is as follows.

It is expressed as Assuming that speech and noise are uncorrelated with each other, the following equation holds.

^{x1 2 = s 2 + n 2} (3) (2), (3) from equation, the power n ² of the noise component contained in x1 is understood to be a quadratic function of the filter characteristics H1～hM. Therefore, the problem of the filter control for minimizing the power n ² of the noise component under the constraint condition becomes a well-known problem of minimizing the quadratic function with the constraint condition.

For various solutions to various constraints and specific algorithms, refer to the literature (“Introduction to Adaptive A
rrays "RAMonzingo et al, John Wiley & Sons, NEW YO
RK, 1980) and U.S. Pat. No. 4,536,887.

As described above, reducing the noise component included in x1 corresponds to reducing the sensitivity of the array system to the direction of arrival of the noise. As a result, the array system has high sensitivity in the target direction, A directional characteristic having low sensitivity is formed in the source direction.

FIG. 5 shows an example of the directional characteristics formed by the adaptive array.
Is shown. In FIG. 5, reference numeral 3 denotes a speaker as in the previous embodiments, and reference numerals 63 and 64 denote noise sources. As can be seen from FIG. 5, the adaptive array does not have sharp directional characteristics, but realizes directional characteristics with low sensitivity in the direction of the noise source. The low sensitivity part of this directional characteristic is called "blind spot", and when the microphone array is composed of M elements, the array system can form M-1 blind spots.

An adaptive array that forms such a directional characteristic has an S / N ratio that is higher than that of a super-directional receiver when a large number of noises reflected indoors come from directions other than the noise source. small. However, regardless of the position of the noise source, almost constant
The feature that an SN ratio can be obtained and the feature that the SN ratio does not fluctuate due to the movement of the speaker 3 because it does not have a sharp directivity in the direction of the speaker 3 use the power difference between the two signals. This is a sound receiver that is very suitable for ensuring the stability of the difference in the S / N ratio required when voice section detection is performed.

In addition, the adaptive array has a feature of reducing the temporal fluctuation of noise power. This will be described with reference to FIGS. 6 (a) and 6 (b). Generally, in a room, noise reflected from walls, floors, ceilings, and the like from other than the direction of the noise source enters the sound receiver. The adaptive array cannot form a blind spot in all of the noise directions, and when the microphone array is composed of M microphone elements, a maximum of M-1 in the direction in which the direct sound and the high-energy reflected sound are incident. The S / N ratio is improved by forming a blind spot.

This effect will be described with reference to FIGS. 6 (a) and 6 (b).
FIG. 6 (a) shows a pulsed noise when a sound is received by an omnidirectional microphone, and FIG. 6 (b) shows a pulsed noise when a sound is received by an adaptive array. In FIG. 6 (a), 71
Is noise directly received from the noise source, and 72, 73, and 74 are noises that have been reflected once or more than once on a wall or floor. Reflected sounds 72, 73, 74 compared to the energy of direct sound 71
Decay 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 the directions of the reflected sounds of 72 and 73. Accordingly, although the power of the reflected sound of the noise indicated by 74 in the adaptive array output shown in FIG. 6 (b) is not much different from that of the sound received by the omnidirectional microphone, the direct sound of the noise and the reflected light of the noise 72 and 73 are not significant. The power of the sound has dropped significantly. As a result, it can be seen that the temporal fluctuation of the noise power is reduced.

As described above, a major factor of erroneous detection of a speech section is a large temporal variation in noise power. To cope with this temporal variation, voice section detection using the power difference between the two signals is performed. However, since it is impossible to completely remove various causes of SN ratio variation, erroneous detection is performed.
100% can not be avoided. Therefore, the feature of the adaptive array used in the present invention, which reduces the temporal fluctuation of the noise power, is very effective in reducing the erroneous detection of the voice section.

FIG. 1 shows a second sound receiver in the configuration example of the present invention.
As the 42, the simplest method is to use one of the microphone elements constituting the microphone array 51. This example is shown in FIG. 7 described later.

As shown in FIG. 10, the second sound receiver inputs some of the microphone outputs of the microphone array 51 of the first sound receiver 42 to the synthesizer 52A, and obtains the output. It is also possible to obtain a second signal x2.

Another example of a microphone array system having a directivity control function is a sound receiving system as shown in US Pat. No. 791,418. In this method, signal processing is performed such that an audio signal having a clear arrival direction is stored and noise arriving from a uniform surrounding area is reduced. In order for this method to work well, the condition that the position of the speaker and the position of the noise source do not match (the direction seen from the microphone may be the same) is necessary, and only the sound from the sound source at the desired position is required. Can be regarded as a kind of directivity control in the sense of extracting

FIG. 7 is a diagram for more specifically explaining the first embodiment of the present invention shown in FIG. In the figure, 51 is a microphone array, 52 is a directivity control unit, 43 is a first short-time power calculation unit, 44 is a second short-time power calculation unit, 45
Is a voice section detection unit based on the power difference, as in the previous embodiments. 81 is a directivity control unit
A first amplifier 82 connected to the output of 52 for receiving the signal x1 and sending the output to the power calculator 43, 82 is a microphone 42
(In this example, one of the microphone elements constituting the microphone array 51 is used). The second amplifier 83 receives the signal x2 and sends the output to the power calculator 44. 83 is the output of the power calculators 43 and 44. A differentiator receiving p1 and p2, 84 is a determining unit that receives the output p1 of the power calculator 43 and is based on short-term section power that may form part of a voice section, and 85 is an output of the differentiator 83 A judgment unit 86 based on the received power and a speech section candidate verification unit or a speech section determination unit 86 that receives the output s1 of the judgment unit 84 based on the short-time power and the output s2 of the judgment unit 85 based on the power.

The procedure for performing this method is as follows. First,
The sound on which the noise is superimposed is received by the microphone array 51. The output signal of the microphone array 51 is input to the directivity control unit 52, and generates a first signal x1. on the other hand,
The output of one microphone element constituting the microphone array 51 is x2. At this time, as a result of the directivity control by the directivity control unit 52, the SN ratio at x1 is larger than the SN ratio at x2.

Next, the levels of the signals are corrected using the amplifiers 81 and 82 so that the powers of the voices included in the signals x1 and x2 become equal. This operation is not essential, but if it is performed, the following description will be simplified. Next, the short-time power calculation units 43 and 44 calculate and output short-time powers P1 and P2 of x1 and x2, respectively. It is assumed that the short-time powers P1 and P2 are represented by logarithmic values (dB) or exact values.

Next, the power P1 of the signal having the high SN ratio is input to the power-based determination unit 84. When the value of P1 is larger than the predetermined threshold Th, the determination unit 84 based on this power outputs “1” as the output S1 to indicate the possibility that the corresponding short time section is a part of the voice section. "" Is output, otherwise "0" is output.

Next, the difference PD between P1 and P2 (PD = P2−
P1) and calculates the difference PD based on the power difference
Enter 5 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 value Pth, and outputs "0" otherwise.

Finally, the output S1 of the determination unit 84 based on the power and the output S2 of the determination unit 85 based on the power difference are input to the voice section determination unit 86. When the values of S1 and S2 are both “1”, the voice section determination unit 86 determines that the short-time section that is a candidate forms a part of a correct voice section, and otherwise determines that the short-time section is a noise section. Is output.

Next, the operation of the voice section detection section 45 based on the power difference will be described with reference to FIGS. 8 (a), (b) and (c). FIG. 8A shows a temporal change of the power P1 at the output of the first sound receiver, and FIG. 8B shows a temporal change of the power P2 at the output of the second sound receiver. And FIG. 8 (c)
The difference PD between P2 and P1 (PD = P2-P1) is shown. In each figure, the vertical axis represents the short-time power of the signal, and the horizontal axis represents time. Reference numeral 11 denotes stationary noise, reference numerals 121 and 122 denote non-stationary noise, and reference numeral 13 denotes voice in the same manner as described in the above example.

Since the powers of the voices included in P1 and P2 are adjusted to be equal, if the power of the stationary noise in P2 is slightly smaller than the power of the voice, the power is indicated by a logarithmic value. In the figures (a) and (b),
The power in the voice section is almost equal. On the other hand, the second
Since the output of the receiver of FIG. 7 has a smaller SN ratio than the output of the first receiver, the noise power in FIG. 8B is smaller than the noise power in FIG. 8A. It is shown that the distance is increased by an amount corresponding to the difference between. As a result, the value of the power difference PD between P2 and P1 shown in FIG. 8 (c) becomes zero in the voice section and takes a non-zero value in the non-voice section.

However, in a real environment, as described above, various SN
Since there is a variation factor of the ratio difference, even in the present invention in which the variation factor is reduced by using a microphone array system having a directivity control function, the PD value assumes such an ideal value. Not necessarily. For example, if the movement of the speaker exceeds the expected range, the value of PD is set to a value larger than zero even in the voice section, and noise coming from the same direction as the voice (for example, the tongue of the speaker, Regarding the sound of the speaker turning over the paper), the value of PD is zero in the noise section even if the power is relatively small.

In consideration of such points, in the present invention, first, as an operation of the determination unit 84 based on the power, as shown in FIG. Would. As a result, for example, even if the noise indicated by the symbol 122 is noise arriving from the same direction as the voice and the PD is small in the noise section, this noise section is not erroneously detected as the voice section. It can be seen that highly effective voice section detection is realized.

As shown in FIG. 1, the voice section determination unit 86 shown in FIG. 7 sets the output s1 from the power-based determination unit 84 to “1” for both the output s2 from the power difference-based determination unit 85. At one time, in addition to the voice section candidate testing section 86a that determines the short-time section as a voice section, only when the time section determined by this testing section to be a voice section continues beyond the predicted value of the minimum duration section of voice Alternatively, a section test unit 86b for determining this time section as a voice section may be provided.

The following experiments were performed to confirm the effectiveness of the present invention.

(Experiment conditions) The experiment was performed in a room where the reverberation time was 0.4 seconds. As noise, a disturbing sound (radio news) was uttered from a speaker. Word sounds (city names) were used as desired sounds, and 100 words uttered under different disturbing sounds were collected. The position of the speaker and the noise source were set 45 degrees apart from the sound receiver. As the sound receiving device 1, an AMNOR sound receiving device (reference: Y. Kaneda and JO), which is one of adaptive arrays
hga "Adaptive Microphone-array System for Noise R
eduction ", IEEE Trans.on Acoust., Speech, Signal Proc
essing, vol. ASSP-34, PP.1391-1400, Dec. 1986). The AMNOR sound receiving device is realized by combining a microphone array composed of a plurality of microphone elements and a digital filter, and is capable of receiving a high SN ratio of about 10 to 16 dB as compared with a single microphone element. As the sound receiver 2, one microphone element which is a component of the microphone array was used. The calculation of the short-time power was performed every 10 ms with a window length of 30 ms.

The threshold value Th in the power-based determination unit 84 is obtained by taking each utterance with a fixed length (1 second), finding the difference PMM between the maximum short-time power and the minimum short-time power in the utterance, Th = PMM ×
0.5. Further, the threshold value Pth of the PD was set to 8 dB.

Note that, as the correct answer of the voice section, a section obtained by applying the first conventional method (a method using only the power-based determination) to the voice without noise was used.

(Experimental Results) Under the above conditions, a word section detection experiment was performed by setting the S / N ratio of the sound at the sound receiving point to be −5 dB at the output of the sound receiver 2.

FIG. 9 shows an example of the experimental result. FIG. 9 (a) shows the speech power and the correct answer in the speech section when there is no noise. Ninth
FIG. 6B shows the output power P2 of the second sound receiver when the disturbing sound is added. FIG. 9 (c) shows the power P1 of the output of the first sound receiver (AMNOR sound receiving device) when the disturbing sound is added, and the selected sound section candidate. The hatched portions indicate erroneously detected speech sections. When comparing FIG. 9 (b) and (c),
(B) The time variation of the noise power indicated by the symbol in the figure is
It can be seen that the output of the adaptive array is small in FIG. That is, the sharp peak of the temporal change of the power is flat.

FIG. 9D shows the result of applying the method of the present invention, which is determined as a word section, by arrows. The non-speech section within 200 ms included in the speech section detected in FIGS. 9C and 9D is regarded as a part of the word section. The portion indicated by hatching is a section where erroneous detection (a speech section is determined as 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, the case where errors at the beginning and end of a word section were detected within 50 ms was regarded as correct, and the correct rate was calculated. AMNO with high SN ratio
When the first conventional method, which is most frequently used in current speech recognition devices, is applied to the output of R, the correct answer rate is 43%. In contrast, in the method of the present invention, 96
% Detection result, and the average detection error at the start and end at that time was about 20 ms. From these results, the effectiveness of the present voice section detection method was confirmed.

When a unidirectional microphone is used as the first sound receiver, for example, as shown in FIG. 2 (a), a straight line connecting the speaker and the microphone is placed around the microphone. When the noise source substantially exists within the range of 90 degrees in the direction of the speaker, the correct answer rate of the word section is about 10%, confirming that the present invention is a high-accuracy acoustic signal detection method. . In the present invention, the above-mentioned ± 96% test result is obtained except for the range of ± 30 ° with respect to the straight line connecting the speaker and the microphone.

For applications where slight performance degradation can be tolerated, a so-called super-directional sound receiver and a sound receiver including a selection filter can also be applied as the first sound receiver of the present invention. FIG. 12 shows an example of the configuration. In FIG. 12, reference numeral 51 denotes a microphone array; 91, an adder for realizing superdirectivity; and 92, a processing filter. As described above, when a super-directional sound receiver is used, the fluctuation of the S / N ratio becomes large in the low frequency range and the high frequency range, so that this processing filter has high sensitivity in a range where the movement of the speaker is expected. This problem is improved by extracting only the low-sensitivity band outside the range. The problem with this method is that the frequency band where the SN ratio fluctuation is small does not necessarily match the voice with the band with high energy, so that the SN ratio of the output of the first sound receiver decreases, and erroneous selection in the voice section candidate is reduced. The point is to increase. On the other hand, the advantage of this method is that the system configuration is simple.

In the present invention, the characteristic inherent to the audio signal is not used at all. However, in order to perform voice section detection, it is very effective to use a determination method utilizing 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 judgment method utilizing the properties of the audio signal. For example, a method is known in which a candidate for a voice section shorter than Tc is determined to be noise by using a predicted value Tc of the minimum duration of a voice signal. Eliminating the influence of pulse noise by combining this determination method is a very effective method in voice section detection. In addition, many other determination methods are known, such as a method of determining that a section in which a signal is non-periodic by using the periodicity of a voice signal is non-voice. In these conventional methods, the section determined as a voice section in the present invention is input, and the section is re-determined, or the final decision of the voice section is made by majority decision of the results of a plurality of determinations including the present invention. It can be easily used in combination with the present invention by a method such as performing.

As described above, the present invention can be combined with many conventionally known speech section detection methods, and as a result, it is possible to realize a great improvement in detection performance according to the purpose of use.

As a first application field of the present invention, there is an application to a speech recognition device as described above. A second application field is an acoustic echo canceller. The acoustic echo canceller is a technique for preventing, for example, in a loudspeaker system or the like, a sound from a receiving speaker wrapping around a transmitting microphone and being received, thereby causing a problem such as howling. The principle of the acoustic echo canceller is to estimate the sound transfer characteristics from the receiving speaker to the transmitting microphone, and subtract the sound component from the receiving speaker from the signal received by the transmitting microphone based on the estimation result. is there. Since the transfer characteristic from the receiving speaker to the transmitting microphone changes with time, it is necessary to continuously perform the estimation. However, when performing the estimation, the condition that the sender is not uttering (otherwise, a large Estimation error occurs). However, the determination of the presence or absence of the speaker's utterance is not always successful, which is one of the current problems of 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 receiving speaker as the unnecessary voice,
At a time when it is determined that the target voice is present in a certain time section, it is considered that the speaker is uttering, and if the operation of estimating the transfer characteristic is stopped, a high-performance It becomes possible to realize an acoustic echo canceller.

A third field of application is in voice storage technology. For example, when digitizing a large amount of continuous uttered voice and recording it on a magnetic disk or the like, information compression technology by voice coding is also important.However, a non-voice section is detected and the section is discarded or the section is cut off. It is also a very important technology to record the information with a particularly low information amount. The present invention is applicable to detection of a non-voice section in such a technique.

Furthermore, since the method of the present invention does not utilize the properties inherent in audio signals, any sound other than audio (for example, music, mechanical sound, impact sound, etc.) can be selected as the sound to be detected. . As a result, the present invention can be applied to various application forms including various kinds of monitoring devices and measuring devices.

[Effects of the Invention] As described above, the method of the present invention provides a method of receiving a sound by a first sound receiver (a microphone array system having a directivity control function) and a second sound receiver installed in the same place. The presence of a desired signal is determined by using the short-time power difference between the extracted signals, enabling the detection of a desired voice section in a non-stationary noise environment, which was not possible with this type of conventional system It is assumed that.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining an embodiment of an audio signal detecting method according to the present invention, and FIG. 2 is a diagram for explaining a problem when a unidirectional microphone and an omnidirectional microphone are used. FIG. 3 is a diagram for explaining a problem when a super-directional sound receiver is used, FIG. 4 is a block diagram showing a specific example of the first sound receiver of FIG. 1, and FIG. Is a diagram showing the directional characteristics of the adaptive array, FIG. 6 is a waveform diagram showing a sound signal waveform of pulsed noise when the omnidirectional microphone and the adaptive array are used, and FIG. 7 is a diagram shown in FIG. FIG. 8 is a block diagram more specifically showing an embodiment of the present invention, FIG. 8 is a graph for explaining the operation of the voice section detecting section shown in FIG.
The figure shows the experimental results confirming the effectiveness of the present invention, FIG.
FIG. 12 is a block diagram showing another embodiment of the present invention,
FIG. 13 is a graph showing a first example of a conventional voice section detection method, FIG. 14 is a diagram showing a microphone installation example for explaining a second example of the conventional voice section detection method, and FIG. FIG. 16 is a graph for explaining an ideal operation of the second conventional method, FIG. 16 is a graph showing a positional relationship between a microphone and a noise source, and FIG.
The figure is a graph for explaining the problem of the second conventional method.
The figure shows the positional relationship between the microphone and the noise source.
FIG. 2 is a block diagram showing a third example of the conventional voice section detection method, and FIG. 20 is a graph for explaining a problem of the third example shown in FIG. 41,41 …… Sound receiver, 43,44 …… Short-time power calculator, 45…
... Sound section detector, 51 ... Microphone array, 52 ...
Directivity control section, 84, 85... Determination section, 86... Voice section determination section.

Claims (7)

(57) [Claims] A first and a second sound receiver which are provided at substantially the same position and have different power ratios (SN ratios) of a target signal and a noise, and have different directional characteristics; When the difference or ratio of the power of the signals transmitted from these sound receivers in a certain time interval is within a predetermined range, it is determined that the target signal has been received in this time interval, One sound receiver comprises an adaptive microphone array system including a microphone array composed of a plurality of microphone elements and a directional characteristic control circuit arranged at a subsequent stage to control the directional characteristic according to the noise position. A method for detecting an acoustic signal, comprising: 2. A signal output from a sound receiver having a high SN ratio, wherein the difference or ratio between the powers of the two signals in a certain time interval is within a predetermined range. If the power in a certain time interval is within a predetermined range, it is determined that the target signal has been received in this time interval. 3. The acoustic signal detection method according to claim 1, wherein said second sound receiver is also constituted by a microphone array system. 4. The apparatus according to claim 1, wherein the time section in which it is determined that the target signal has been received continues beyond the predicted value of the minimum duration of the voice, and the target signal is received in this time section. A sound signal detection method characterized by determining. 5. The device according to claim 1, wherein the second sound receiver is
An acoustic signal detection method comprising using one microphone element which is a component of a microphone array constituting the first sound receiver. 6. The apparatus according to claim 4, wherein the second sound receiver comprises:
An acoustic signal detection method, characterized in that some of the microphone elements constituting the first sound receiver are shared, and further comprising means for synthesizing the outputs of the several microphone elements. 7. A first and a second sound receiver provided at substantially the same position and having different power ratios (SN ratios) of a target signal and a noise and having different directional characteristics are used, When the difference or ratio of the power of the signals transmitted from these sound receivers in a certain time interval is within a predetermined range, it is determined that the target signal has been received in this time interval, One sound receiver includes a directional microphone array in which a plurality of microphones are arranged, a synthesizer that receives the output of each microphone to synthesize superdirectivity, and receives a predetermined band component by receiving the output of the synthesizer. An acoustic signal detection method, comprising: a band selection filter to be passed.