JP2982766B2 - Sound source direction estimation method and apparatus - 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 and apparatus, and more particularly, to a sound source direction estimating method for estimating a sound source direction by a cross-correlation function calculated using signals received by a plurality of microphones. And its device.

[0002]

2. Description of the Related Art FIG. 3 shows "Journal of the Acoustic Society of July, 1992.
Japan, Vol. 13, No. 4, pp. 241-252 (The Journal of the Acoustic)
Ial Society of Japan, vol.
13, no. 4, pp. 241-252, July 1
FIG. 992) is a block diagram of a conventional sound source direction estimating apparatus (hereinafter, referred to as a first conventional example).

[0003] In FIG. 3, signals input from microphones 101 and 102 from a sound source 115 are amplified by amplifiers (hereinafter abbreviated as amplifiers) 103 and 104 and A / D
The signals are converted into discrete signals by converters 105 and 106. Auto-regression filter coefficients are obtained by auto-regression coefficient calculation sections 109 and 110 using discrete signals in which high-frequency components are emphasized by pre-emphasis processing sections 107 and 108. The filter coefficients and the discrete signal are convolved by the inverse filter processing units 111 and 112. Outputs of the inverse filter processing units 111 and 112 are whitened signals. The cross-correlation function for these whitening signals is obtained by the cross-correlation function calculation unit 113, and the direction corresponding to the time difference giving the maximum value of the cross-correlation function is estimated by the direction detection unit 114 as the sound source direction.

FIG. 4 shows “September 1996, IEICE Technical Report, DSP96-77, DSP96-52, SP96-52,
It is a block diagram of a sound source direction estimating device (hereinafter, a second conventional example) described on page 3-29.

In FIG. 4, signals input from a sound source 128 to microphones 116 and 117 are amplified by amplifiers 118 and 1.
The signal is amplified at 19 and converted to discrete signals by A / D converters 120 and 121. The discrete signal has a zero-crossing which is 1 when the signal level becomes 0 in the zero-crossing time series generation units 122 and 123, -1 when the signal level is negative, and 0 when the signal level is negative. Converted to time series.

The cross-correlation function calculator 124 calculates the cross-correlation function of these zero-crossing time series, and the normalized power calculator 1
At 25, the power is normalized. Time average calculator 126
Then, this normalized power value is time-averaged, and the direction detection unit 1
At 27, the direction corresponding to the time difference giving the maximum value is estimated as the sound source direction.

[0007]

However, the prior art described above has the following drawbacks.

The first problem is that the calculation accuracy required for the cross-correlation calculation in the first conventional example is high and the hardware scale is large.

The reason is that the A / D converters 105, 1
This is because the cross-correlation function is calculated using the multi-bit discrete signal quantized according to 06 as it is.

A second problem is that the first conventional example is susceptible to noise.

The reason is that the sound source direction estimation is performed by using the cross-correlation function calculated by the cross-correlation function calculation unit 113 by the direction detection unit 114 without time averaging.

A third problem is that the accuracy of estimating the sound source direction is low in the second conventional example.

The reason is that the zero-crossing time series generation unit 12
This is because the zero-crossing time sequence generated in 2, 123 has a value only at the zero-crossing point of the input signal, and the input signal samples are equivalently thinned out. That is, the resolution decreases.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned conventional circumstances, and has been made to solve the above-mentioned drawbacks inherent in the conventional technology. It is an object of the present invention to provide a novel sound source direction estimating method and a device thereof having a small wear scale.

It is another object of the present invention to provide a novel sound source direction estimating method and apparatus having excellent noise immunity.

Still another object of the present invention is to provide a novel sound source direction estimating method and apparatus capable of performing highly accurate sound source direction estimation.

[0017]

In order to achieve the above object, a sound source direction estimating apparatus according to the present invention receives a first audio signal as an input and outputs "1" when the signal level is positive and when the signal level is negative, A first code time sequence generation unit (7 in FIG. 1) for generating a code time sequence directly represented by “0” when “−1” or 0, and a signal level of the second audio signal as an input. Is "1" when is positive and "-" when it is negative.
A second code time sequence generation unit (8 in FIG. 1) that generates a code time sequence directly represented by “0” when the value is “1” or “0”;
A first cross-correlation function calculator (9 in FIG. 1) for calculating a correlation using an output of the first code time sequence generator and an output of the second code time sequence generator; A first normalized power calculator (10 in FIG. 1) that normalizes the output of the cross-correlation function calculator, and a first time average calculator (10 in FIG. 1) that averages the output of the first normalized power calculator. 1) and a first direction detection unit (12 in FIG. 1) for estimating the sound source direction using the output of the first time average calculation unit.

[0018]

According to the present invention, after the input signal is whitened, "+
It is converted into a code time sequence represented by a ternary value of “1”, “−1”, or “0.” For this reason, necessary calculation accuracy is reduced and the hardware scale can be reduced.

In the present invention, the direction of the sound source is estimated by time-averaging the normalized power value of the cross-correlation function. For this reason, the direction estimation is not mistaken due to the influence of noise.

Further, in the present invention, direction estimation is performed using a code time sequence instead of a zero crossing time sequence. Therefore, the same number of signal samples as the input signal can be used, and there is no deterioration in estimation accuracy due to thinning.

[0021]

Next, preferred embodiments of the present invention will be described in detail with reference to the drawings.

[Description of Configuration] FIG. 1 shows a first embodiment of the present invention.
FIG. 2 is a block diagram showing an embodiment.

Referring to FIG. 1, a signal X arriving from a sound source 13 whose direction is to be estimated in microphones 1 and 2 is shown.
₀ (t) and Y ₀ (t) are obtained as input signals. These input signals are amplified by amplifiers (hereinafter abbreviated as amplifiers) 3 and 4, respectively, quantized by A / D converters 5 and 6, and converted into discrete signals. This discrete signal is “1” when the signal level is positive, “−1” when the signal level is negative,
When it is 0, it is converted by the code time sequence generation units 7 and 8 into a code time sequence represented as “0” as it is. Using these code time sequences, a cross-correlation function is calculated by a cross-correlation function calculator 9.
Ask for.

The cross-correlation function obtained by the cross-correlation function calculator 9 is converted into a power value normalized by the normalized power calculator 10 and time-averaged by the time average calculator 11 of the normalized power. . In the direction detection unit 12, the direction corresponding to the time difference that gives the maximum value of the time-averaged normalized power is estimated as the sound source direction.

[Description of Operation] Next, the operation of the present invention will be described with reference to FIG.

In FIG. 1, a signal input from the sound source 13 to the microphone 1 is X ₀ (t), and a signal input from the sound source 13 to the microphone 2 is Y ₀ (t). The input signal X ₀ (t) is amplified by the amplifier 3, quantized by the A / D converter 5, and converted into a discrete signal X ₁ (n).
The input signal Y ₀ (t) is amplified by the amplifier 4, quantized by the A / D converter 6, and converted into a discrete signal Y ₁ (n).

Next, the discrete time signal X ₁ (n) is “1” when the signal level is positive and “−” when the signal level is negative in the code time sequence generation unit 7.
When it is 1 or 0, it is directly converted into a code time sequence x (n) represented by “0.” The code time sequence x (n) for the discrete signal X ₁ (n) is defined by Expression 1.

[Equation 1]

[0029]

The discrete signal Y ₁ (n) is also converted into a code time sequence by the code time sequence generator 8 in the same manner as X ₁ (n). Code time sequence y for discrete signal Y ₁ (n)
(N) is defined according to Equation 1. The cross-correlation function calculation unit 9 calculates the cross-correlation function Rxy (k) for the code time series x (n), y (n) using Equation 2.

[Equation 2]

[0032]

The incident direction of the sound wave to be originally sought (sound source direction)
θ (rad) can be obtained by Expression 3.

[Equation 3] θ = sin ⁻¹ (τc / M) Here, τ (sec) is the time when the signal X ₀ (t) reaches the microphone 1 and the signal Y ₀ (t) reaches the microphone 2 M (m) is the interval between two microphones, and c (m / sec) is the speed of sound. Actually, the code time sequence x (n), y of the sampling frequency fs (Hz)
To calculate the correlation using (n), the time difference between the signals received by the two microphones is calculated by the corresponding number of samples k
It is represented by Here, k is an integer. Therefore, the time difference τ (sec) actually obtained is as shown in Expression 4.

[Equation 4] τ = k / fs By substituting Equation 4 into Equation 3, Equation 5 is obtained.

[Equation 5] θ = sin ⁻¹ ((c / M) × (k / fs)) In Expression 5, θ is represented by [rad].
g], Expression 6 obtained by multiplying Expression 5 by 180 / π must be used.

[Equation 6] θ = (180 / π) sin ⁻¹ ((c / M) × (k / f)
s)) On the other hand, the time difference τ (sec) is maximized when the sound source is located right beside (90 degrees) which is an extension of two microphones arranged on a straight line. Therefore, the number of samples k _max corresponding to the maximum value can be obtained by Expression 7.

[Equation 7] k _max ≦ fs × (M / c) That is, when a discrete signal is input, the maximum value of the sound source direction angle that can be detected is generally smaller than 90 degrees. When the arrival direction of the sound source is equal to or larger than the angle corresponding to k _max , the sound source direction cannot be directly obtained by Expression 6. So, k
_The cross-correlation function is obtained up to _max + 1 and _kmax + 2, and when the peak value exists at _kmax + 1, it is determined that the arrival direction is 90 degrees.

Next, the normalized power calculation unit 10 calculates the normalized power p (k) from the cross-correlation function Rxy (k) using Expression 8.
Ask for.

[Equation 8]

[0041]

The power p (k) obtained from Equation 8 is represented by code time series x (n), y (n) (n = 0, 1,
2,. . . . . , N−1), but N cannot be made large in consideration of the unsteadiness of the audio signal. Therefore, when the total data length of voice is L, 1
Assuming that a frame is N samples, u samples are overlapped and divided into F frames, and F frames corresponding to each frame are divided.
Pieces of normalized power pf (k) (f = 0, 1,..., F
Use the average of -1) instead of p (k). The time average calculator 11 calculates the average value of P _f (k) C _p (k) (k =
.., K _max , k _max +1 and k _max +
2) is obtained by Expression 9.

[Equation 9]

[0044]

P _f (k) represents the power of the f-th frame,
x (n), y (n) (n = f (Nu), f (Nu)
+1, f (N−u) +2,. . . . , F (N−u) + N
-1) is obtained using Expressions 7 and 8. However,
F is the total number of frames in the averaging (the number of analysis sections), N-
u is a code time sequence x newly input in each frame.
(N) and y (n) represent the number of samples.

Finally, the direction detector 12 calculates the time average C _p
K1 that gives the maximum value of (k) is obtained, and an estimated angle θ degree corresponding to Expression 10 is obtained.

[Equation 10]

[0048]

[0049]

FIG. 2 is a block diagram showing a second embodiment according to the present invention.

Referring to FIG. 2, the difference between the second embodiment and the first embodiment shown in FIG.
The A / D converter 1 is added to the code time sequence generation units 24 and 25.
That is, the signals supplied from 8 and 19 are processed by an inverse filter using an auto-regression coefficient to whiten. To this end, auto-regression coefficient calculation units 20 and 21 and inverse filter processing units 22 and 23 are provided. However, other configurations / operations are the same as those in FIG.

Using the discrete signals X ₁ (n) and Y ₁ (n), filter coefficients for the inverse filter are obtained by the auto-regression coefficient calculators 20 and 21. The discrete signals corresponding to these filter coefficients are convolved by the inverse filter processing units 22 and 23 to generate a whitened signal.

Next, using the discrete signal X ₁ (n), an auto-regression coefficient calculation section 20 obtains an auto-regression coefficient α _k (k = 1, 2,..., Q). The auto-regression coefficient is Le
As known as the Vinson-Durbin method, it can be obtained by using Expression 11 representing the autocorrelation function of the input signal. For the Levinson-Durbin method, see "Proceedings
Of the iEiEi, Vol. 63, No. 4
Issue, pages 561-580 (Proceedings
of the IEEE, vol. 63, No. 4, p
561-580, April, 1975).

The autocorrelation function Rxx (k) (k
= 0, 1, 2,. . . . , Q-1) to determine the autoregressive coefficient α _k ^(Q) . Here, Q is the order of the autoregressive coefficient, σ ^{q 2} is the variance, and C _q is the reflection coefficient.

[Equation 11]

[0055]

The initial value is set as shown in Expression 12.

[Equation 12] σ ₀ ² = Rxx (0) α ₀ ⁽⁰⁾ = 1 α ₁ ⁽¹⁾ = C ₁ = −Rxx (1) / σ ₀ ² σ ₁ ² = σ ₀ ² (1- C ₁ ² ) q = 1 Next, the variable q is updated by the recurrence formula 13, and the Q-order autoregressive coefficient α _k ^(Q) to be obtained is sequentially obtained.

[Equation 13]

[0059]

In the same manner as the auto-regression coefficient calculator 20, the auto-regression coefficient β _k (k) for the discrete signal Y ₁ (n) is calculated.
= 1, 2,. . . . , Q) are calculated by the auto-regression coefficient calculator 21. The autoregression coefficient α _k (k = 1, 2,..., Q) obtained by the autoregression coefficient calculation unit 20 and the discrete signal X
₁ (n) is convolved by the inverse filter processing unit 22 using Expression 14, and a whitening signal e _X (n) is obtained.

[Equation 14]

[0062]

The self-regression coefficient β _k (k = 1, 2,..., Q) obtained by the self-regression coefficient calculation section 21 and the discrete signal Y ₁ (n) are The whitening signal e _Y (n) is obtained by convolution with the same calculation method as in Expression 14. The obtained whitening signals e _X (n) and e _Y (n) are converted into code time sequences x (n), y by code time sequence generation units 24 and 25.
By converting to (n), sound source direction estimation can be performed in the same procedure as in FIG.

[0064]

The present invention is configured as described above,
According to the present invention, the following effects can be obtained.

The first effect is that the calculation accuracy required for calculating the cross-correlation function is reduced, and the hardware scale can be reduced.

The reason is that the signal time is converted into a code time series that is "1" when the signal level is positive, "-1" when the signal level is negative, and "0" when it is 0.

The second effect is excellent in noise resistance.

The reason is that the power values of the cross-correlation function are averaged over time.

The third effect is that a highly accurate direction estimation of a sound source can be performed.

The reason is that the input signal is whitened.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment according to the present invention.

FIG. 2 is a block diagram showing a second embodiment according to the present invention.

FIG. 3 is a block diagram showing a first example of a conventional sound source direction estimating apparatus.

FIG. 4 is a block diagram showing a second example of a conventional sound source direction estimation device.

[Explanation of symbols]

 1, 2, 14, 15 ... microphone 3, 4, 16, 17 ... amplifier 5, 6, 18, 19 ... A / D converter 7, 8, 24, 25 ... code time sequence generation unit 9, 26 ... cross-correlation function Calculation units 10, 27 Normalized power calculation unit 11, 28 Time average calculation unit 12, 29 Direction detection unit 13, 30 Sound source 20, 21 Autoregression coefficient calculation unit 22, 23 Inverse filter processing unit

Continued on the front page (58) Fields surveyed (Int.Cl. ⁶ , DB name) G01S 3/80-3/86 G01S 5/18-5/30 G01S 7/52-7/64 G01S 15/00-15 / 96 G06F 15/336

Claims (5)

(57) [Claims] 1. A method for extracting the polarities of a plurality of audio signals obtained by a plurality of microphones, and
Generating a plurality of code time sequences corresponding to one to one, calculating a cross-correlation of the plurality of code time sequences, calculating a normalized power of the cross-correlation, calculating a time average of the normalized power, A sound source direction estimating method characterized by estimating a sound source direction by time averaging. 2. The code time sequence according to claim 1, wherein the code time sequence is expressed as “1” when the signal level is positive, “−1” when the signal level is negative, and “0” as it is when the signal level is 0. The described sound source direction estimation method. 3. A method for calculating a plurality of auto-regression coefficients corresponding to a plurality of audio signals acquired by the plurality of microphones in a one-to-one manner, and using the plurality of auto-regression coefficients as coefficients to inverse the plurality of audio signals. Filter,
Extracting the polarities of the outputs of the plurality of inverse filter processes to generate a plurality of code time sequences, calculating a cross-correlation of the plurality of code time sequences, calculating a normalized power of the cross-correlation,
A sound source direction estimating method, comprising calculating a time average of the normalized power and estimating a sound source direction by the time average. 4. A plurality of code time sequence generators for extracting a plurality of audio signals obtained by a plurality of microphones and a polarity of the plurality of audio signals in a one-to-one correspondence, and the plurality of code time sequences. A cross-correlation function calculator that receives all outputs of the generator and calculates a cross-correlation, a normalized power calculator that receives outputs of the cross-correlation function calculator and calculates a normalized power, and the normalized power calculator A sound source direction estimating apparatus, comprising: a time average calculating unit that calculates the time average of the power value in response to the output of the sound source direction; and a direction detecting unit that estimates the sound source direction based on the output of the time average calculating unit. 5. A plurality of auto-regression coefficient calculators for calculating a plurality of audio signals acquired by a plurality of microphones and auto-regression coefficients thereof in one-to-one correspondence with the plurality of audio signals; A plurality of inverse filter processing units for performing an inverse filter process on the plurality of audio signals by using an output of the regression coefficient calculation unit as a coefficient; and a plurality of code time sequence generating units for extracting polarities of outputs of the plurality of inverse filter process units , A cross-correlation function calculator that receives all outputs of the plurality of code time sequence generators and calculates a cross-correlation, and a normalized power calculator that receives output of the cross-correlation function calculator and calculates normalized power Unit, a time average calculation unit that receives the output of the normalized power calculation unit and calculates a time average,
A sound source direction estimating device, comprising: a direction detecting unit that estimates a sound source direction based on an output of the time average calculating unit.