JP2985982B2 - Sound source direction estimation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sound source direction estimating method for estimating the direction and position of a sound source based on a cross-correlation function between signals observed by a plurality of microphones.

[0002]

2. Description of the Related Art The direction of a sound source is estimated, for example, by estimating the direction in which an abnormal sound is generated by a remote monitoring device or the like, and pointing a monitor camera in the direction, or aiming the camera in the direction of a speaker in a video conference. It is a basic technology required for various purposes such as tracking the flight trajectory of an aircraft outdoors.

The most basic conventional method of sound source direction estimation is
This is a method based on a time difference between a plurality of signals observed by a plurality of microphones. FIG. 7A is a diagram for explaining this. For the first microphone 1 and the second microphone 2,
The first microphone 1 is incident on the wavefront 3 from the incident direction 4.
Let the output signal 5 of the second microphone 2 be x (t) and the output signal 6 of the second microphone 2 be y (t). In the state shown in this figure, the sound wave is first received by the first microphone 1 and is received by the second microphone 2 with a slight delay. Assuming that the distance between the first and second microphones 1 and 2 is d and the arrival direction of the sound wave is an angle θ shown in the drawing, the delay time (time difference) of arrival at the second microphone 2 is the distance The time required to advance d · sin θ is τ0, and when the sound speed is represented by c, τ0 = (d · sin θ) / c (1) Assuming that the arrival direction of the sound wave is one direction, the output signal y (t) of the second microphone 2 is expressed by y using the output signal x (t) of the first microphone 1.
(T) = x (t−τ0). And
If τ0 can be obtained from these two signals x (t) and y (t), from the equation (1), the arrival direction of the sound wave is given by the following equation: θ = sin ⁻¹ (c · τ0 / d) (2) You can ask.

[0004] The time difference τ0 of the sound wave is represented by two signals x
The cross-correlation function φxy (τ) of (t) and y (t) is calculated, and can be obtained as the value of τ that takes the maximum value. Here, the cross-correlation function of the discretized signal (t is an integer) is defined by the following equation, and φxy (τ) = {x (t) y (t + τ) (3)} is defined by addition of t. In that case, the sum (Σ) is changed to integral). At this time, if the relation y (t) = x (t−τ0) is used, φxy (τ) = {x (t) x (t + τ−τ0) = φxx (τ−τ0) (4)} Addition. Here, φxx (τ) is an autocorrelation function of x (t), and takes a maximum value when τ = 0 as is known. Therefore, it is understood that φxy (τ) takes the maximum value when τ = τ0. FIG. 7B illustrates signals x (t) and y (t) and a cross-correlation function φxy (τ) calculated from the signals x (t) and y (t), taking a case where the sound wave is a pulse sound as an example. φxy
It can be seen that (τ) has a clear maximum at τ = τ0.

[0005]

In this conventional method, a sound wave other than a sound wave generated by a sound source to be estimated exists, and even if a reflected sound exists, the power of the sound is small. Is known to operate well (takes the maximum value at about τ = τ0). However, especially when considering the estimation of the sound source direction in a room sound field, the power of the reflected sound is often large, which greatly affects the conventional method. This will be described with reference to FIG. 7C.

FIG. 7C illustrates the signals x (t) and y (t) and the cross-correlation function φxy (τ) calculated from them when the sound wave is a pulse sound and there is a single reflected sound as an example. The reflected sound 8 is received for the direct sound 7. The direct sound 7 means a sound that reaches the microphone directly from the sound source, and its arrival direction matches the direction of the sound source.
On the other hand, the reflected sound 8 is a sound that is emitted from a sound source and is reflected by a wall or the like and reaches the microphone. Therefore, the arrival direction of the reflected sound is generally different from the sound source direction. Therefore, only the arrival time difference of the direct sound 7 includes information on the sound source direction.

In the example of FIG. 7C, the direct sound 7 and the reflected sound 8 are included in the two microphone output signals x (t) and y (t) with different time differences τ0 and τ1, respectively. Only the time difference τ0 of the direct sound 7 includes the information of the sound source direction. At this time, the signal x (t),
The cross-correlation function φxy (τ) calculated from y (t) is as shown in the figure, and as can be seen by comparing FIG. 7B, the local maximum occurs at a plurality of points other than τ = τ0 and the maximum value It can be seen that the value of τ that gives Further, since the number of points giving the maximum value increases in proportion to the square of the number of reflected sounds, in a room sound field where there are many reflected sounds having a large power, the value of τ0 is obtained by a conventional method to determine the value of τ0. It can be seen that it is difficult to estimate the direction.

An object of the present invention is to solve the above-mentioned problems of the conventional sound source direction estimating method, and to realize a novel sound source direction estimating method which realizes a good sound source direction estimation even in a room sound field having many reflected sounds. Is to provide.

[0009]

A method for estimating a sound source direction based on a cross-correlation function between signals observed by a plurality of microphones. It is characterized in that peak hold processing is performed for each. In this way, the influence of the reflected sound is masked, and the time difference estimation of the direct sound is favorably performed.

According to the second aspect of the present invention, the output of each microphone is divided into a plurality of frequency components, and each of the divided components is subjected to a peak hold process. A cross-correlation function is calculated, these cross-correlation functions are weighted and averaged, and the sound source direction is estimated based on the value. According to the third aspect of the present invention, the peak-hold processing in the first or second aspect is logarithmized, and the cross-correlation function is obtained for the logarithmized processing.

[0011]

[Operation] First, the operation and effect of the peak hold processing will be described. The information on the sound source direction is included only in the time difference τ0 of the direct sound. However, as shown in FIG. 7C, when a reflected sound having a time difference τ1 (≠ τ0) such as the signals x (t) and y (t) is added, the cross-correlation function φxy (τ)
Has no definite maximum at the point τ0. Therefore,
If peak hold processing (processing of holding and outputting the maximum value of the input signal power up to each time point) is performed on the signals x (t) and y (t) shown in FIG. 7C, the result is as follows. The signals are x (t) and y (t) as shown in FIG. 1A. 1A, the signals x (t) and y (t) are held at their peak values when the direct sound 7 is received, and the reflected sound 8 arriving later has a lower power than the direct sound 7 so that the reflected sound 8 Is masked and cannot be observed. These signals x (t) and y (t) that have been subjected to the peak hold processing
X (t) ← x (t) −x (t−1) (5) (or differential processing), only the part where the waveform changes (increases) is extracted, and the difference processing result Are the same signals as the signals x (t) and y (t) when there is no reflected sound shown in FIG. 7B. Accordingly, the cross-correlation function calculated from these signals is the same as φxy (τ) shown in FIG. 7B, and has a clear maximum value at τ0.

Next, the operation and effect of the logarithmic processing will be described. The reflected sound has a longer path from the sound source to the microphone than the direct sound, and the power is smaller than that of the direct sound due to the wall absorption at the time of reflection. Therefore, the above-described peak hold processing is effective. However, in an actual room reflected sound sequence, a plurality of initial reflected sounds having a large power arrive at almost the same time, and as a result, a reflected sound often having a larger power than a direct sound (to be more precise, a plurality of reflected sounds are superimposed). ) Are observed. FIG. 1B shows the result of the peak hold processing when a high-power reflected sound arrives. As can be seen from the figure, such an influence 9 of the reflected sound cannot be removed only by the peak hold processing. However, since the power of the reflected sound is at most several times the direct power, this effect is reduced by logarithmic processing. For example, it is assumed that the power of the background noise (stationary noise) is 1, a direct sound having a power of 1000 arrives there, and a reflected sound having a power of 2000 subsequently arrives. Considering this as a numerical value after logarithmic processing, the background noise is 0 dB, the direct sound has a power of 30 dB, and the reflected sound has a power of 33 dB. Therefore, in the case of an exact value, the magnitude of the reflected sound is twice that of the direct sound, but the result of logarithmic processing of this is 1.1 times, and the effect of the reflected sound is numerically reduced. Recognize. FIG. 1C shows the result of performing logarithmic processing on the signal of FIG. 1B. Effect of reflected sound 9 shown in FIG. 1C
Is smaller than that shown in FIG. 1B, and the effectiveness of the logarithmic processing can be understood.

The band division processing is effective in increasing the chance of observing a direct sound. FIG. 2A is an example of the time-frequency spectrum of the audio signal displayed as contour lines. The time 11 at which the frequency component of 1 kHz to 2 kHz of the sound occurs is slightly behind the time 10 at which the sound occurs. As described above, when an audio signal or the like is divided into frequency components and observed, it is understood that the occurrence time differs for each frequency component. This 1kHz ~ 2k
Although the information of the direct sound is also included in the occurrence part of the Hz component, when the signal is viewed in the entire band, it is masked by the component of 0 to 1 kHz or the like, and good direct sound cannot be observed. For these reasons, performing the signal processing for each band by performing the band division processing improves the accuracy of estimating the time difference τ0 from the viewpoint of increasing the chance of direct sound observation. The invention of claim 2 utilizes this idea.

As described above, the method of the present invention is characterized in that the influence of the reflected sound is removed by peak hold processing and logarithmic processing. As a result, the influence of the reflected sound on the cross-correlation function, which is a problem of the conventional sound source direction estimation method based on the cross-correlation function, is greatly improved.

[0015]

FIG. 3 shows an embodiment of the present invention. The output signals of the two microphones 21 are divided into M frequency bands in each of the two band division units 22. As a method of this band division, for example, FFT (Fast Fourier Transform)
And so on. Next, the following processing is performed on the signals of each band of the two systems. First, the power of the signal is obtained in the power calculator 23. Next, the peak hold processing section 24 performs peak hold processing of the signal. Next, the logarithmic processing unit 25 performs logarithmization of the signal. Next, the time difference processing unit 26 calculates a time difference between the signals. Next, the cross-correlation function calculator 27 obtains a cross-correlation function between signals in the corresponding frequency bands of the two microphone outputs subjected to the same processing. When the signals subjected to the time difference processing with respect to the k-th frequency band are expressed as x ′ (k, t) and y ′ (k, t), respectively, the cross-correlation function φxy ′ (k, τ) becomes Is obtained by the following equation.

Φxy ′ (k, τ) = {x ′ (k, t) y ′ (k, t + τ) (6)} is Performed on t In the sound source direction estimating unit 28, the cross-correlation calculated for each band The function is averaged as in the following equation to calculate φxy (τ). φxy (τ) = {Wkφxy (k, τ) (7)} is from k = 1 to k = M.

Here, Wk is a weighting function assigned to each band at the time of averaging.
This is a value determined by measurement conditions, such as reducing the value of k. The sound source direction estimating unit 28 obtains the value of τ at which the cross-correlation function φxy (τ) obtained by the above operation takes the maximum value, and uses it as the estimated value of the time difference τ0 of the direct sound. Then, the sound source direction is obtained by the following equation θ = sin ⁻¹ (c · τ0 / d) (8) and output.

In the above example, two microphones have been described. However, if three or more microphones can be used, the estimation accuracy is improved. In that case, a plurality of pairs of two microphones are selected from a plurality of microphones, and the sound source direction θ is determined for each pair in the same manner as described above.
Is estimated, and a plurality of obtained sound source directions are averaged to obtain a final estimation result. By averaging a plurality of estimation results in this way, the effects of estimation errors due to noise and the like are reduced.

In FIG. 3, a peak hold process and a logarithmic process may be performed to obtain a cross-correlation function without band division. In any case, the logarithmic processing may be omitted. An acoustic signal generated from a target sound source is often a non-stationary signal such as voice. If an unsteady signal occurs intermittently, the chance of observing a direct sound increases at each occurrence time, so that it is effective to provide the peak hold characteristic with an attenuation characteristic. This will be described with reference to FIGS. 2A, 2B, and 2C. FIG. 2B
Of (a) intended to pulse sound representing the received sound signal of the microphone in the case that occurs intermittently, direct sound 7 _{1-7 3}
Are successively separated in time, and each of these direct sounds 7 _{1 to} 7
Each to ₃ multiple reflections series 8 _{1-8 3} occurs. FIG. 2B shows the result of performing a peak hold process having no attenuation characteristic on this signal. According to this figure, the peak value of the first direct sound 7 ₁ is maintained and the reflected sound is removed, but the second and third direct sounds 7 ₂ and 7 ₃ are also removed. I understand.
Therefore, a peak hold process having an attenuation characteristic is performed on the signal of FIG. The result is shown in FIG. From this figure, if no attenuation to the peak hold, second, third direct sound 7 _2, 7 ₃ without being removed, can be removed reflected sound. In this way, the direct sound observation opportunity is increased, and the time difference τ
Estimating 0 is effective in that it is less susceptible to calculation errors and noises than in the case of estimating τ0 from only one direct sound, and that higher estimation accuracy can be obtained.

[0020]

Next, a description will be given of the results of an experiment conducted by the embodiment of FIG. 3 to verify the effectiveness of the method of the present invention. The experiment was performed in a room having a volume of 50 m ³ and a reverberation time of 0.25 seconds. FIG. 4 shows the arrangement of the sound source and the microphone in the experiment. In the figure, reference numerals 31 and 32 denote sound sources, 33 and 3
Reference numeral 4 denotes a microphone. The signal for sound source direction estimation used speech. The received signal is sampled at 8 kHz, and band-divided (ie, 32
Band), and two seconds of received sound data were used for calculating the correlation.

FIGS. 5A and 5B show experimental results when only the sound source 32 is present. The experimental results indicate the obtained cross-correlation function φxy (τ). FIG. 5A shows the result obtained by the conventional method without performing the peak hold process.
FIG. 5B shows the result obtained by the method of the present invention. The correct time difference corresponding to the direction of the sound source is indicated by a black arrow in the figure. As can be seen from the figure, in the results obtained using the conventional method (FIG. 5 (a)), the maximum value coincides with the correct answer, but the maximum value which does not coincide with the correct answer also occurs due to the influence of the reflected sound. ing. On the other hand, the result ((b) of FIG. 5) using the method of the present invention shows that a clear maximum value is obtained at the correct answer position.

FIGS. 6A and 6B show two sound sources 31 and 32.
The results when a different voice (power ratio 1: 2) is generated are shown. FIG. 6A shows the result obtained by the conventional method, and FIG. 6B shows the result obtained by the method of the present invention. From the figure, it is difficult to distinguish between these two sound sources in the result obtained by using the conventional method (FIG. 6A), but in the result obtained by using the method of the present invention (FIG. 6B). It can be seen that the time difference between direct sound and each sound source can be clearly estimated.

As described above and confirmed by experiments, the present invention is a very effective technique for estimating the direction of a sound source in a sound field where many reflected sounds exist.

[Brief description of the drawings]

1A is a diagram illustrating a signal obtained by performing a peak hold process on the signal of FIG. 7C, and FIG. 1B is a diagram illustrating a signal obtained by performing a peak hold process on a signal in which the power of the reflected sound is greater than the power of the direct sound. FIG. 7C is a diagram illustrating a signal obtained by performing logarithmic processing on the signal of FIG. 1B.

FIG. 2A is a diagram illustrating an example of a time-frequency spectrum of voice, and FIG. 2B is a diagram illustrating the effectiveness of a peak hold process having an attenuation characteristic.

FIG. 3 is a block diagram showing an embodiment of the present invention.

FIG. 4 is a diagram showing experimental conditions for confirming the effectiveness of the present invention.

FIG. 5 is a view showing an experimental result when one sound source is used.

FIG. 6 is a view showing an experimental result in a case where there are two sound sources.

7A is a diagram illustrating the relationship between the arrival direction θ of a sound wave and signals received by two microphones, and FIG. 7B is a diagram illustrating output signals of two microphones when there is no reflected sound and a correlation calculated from them. FIG. 7C is a diagram illustrating the function, and FIG. 7C is a diagram illustrating the output signals of the two microphones when there is a reflected sound and the correlation function calculated from the output signals.

Continuation of the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G01S 3/00-3/86 G01H 3/00

Claims (3)

(57) [Claims] 1. A method for estimating a sound source direction based on a cross-correlation function between signals observed by a plurality of microphones, wherein a plurality of signals are generated by performing peak hold processing on outputs of the plurality of microphones. A sound source direction estimating method, comprising calculating a cross-correlation function from a plurality of processed signals. 2. A method for estimating a sound source direction based on a cross-correlation function between signals observed by a plurality of microphones, comprising: dividing an output of each of a plurality of microphones into a plurality of frequency components; Generating a plurality of signals by performing peak hold processing, calculating a cross-correlation function between corresponding frequency components of the processed plurality of signals, and weighing and averaging the cross-correlation function between these frequency components. A sound source direction estimation method characterized by estimating a sound source direction based on a cross-correlation function obtained by: 3. The sound source direction estimating method according to claim 1, wherein the cross-correlation function is calculated for a logarithmically processed one after the peak hold processing.