TW201101852A - Sound source direction detecting method and apparatus thereof - Google Patents

Abstract

A method for detecting the direction of a sound source is provided. Sound signals from the sound source are received via a microphone array, and then digitized separately for each microphone. By calculating spectral entropy and signal-to-noise ratio, voice activity is detected for each digital signal stream, so as to extract the signal frames carrying speech signals. When a frame of speech signal is detected, time delay of arrival (TDOA) between each pair of microphone is estimated according to the principle of generalized cross correlation function. Then, the sound source's direction is computed to obtain an azimuth and an elevation according to the estimated TDOA vector.

Description

201101852 VI. Description of the Invention: [Technical Field] The present invention relates to a sound source detecting device, and more particularly to a sound source detecting device suitable for detecting a human voice position. [Prior Art] Sound source position detection has a wide range of applications, such as sonar detection, wireless communication, video conferencing systems, and the like. In recent years, with the development of robots, 越来越 has become more and more vigorous. Research on the detection of sound source orientation applied to robots has gradually increased. In addition, the sound source azimuth detection system can also be used to enhance human-robot interaction (HRI), or combine with Automatic Speech Recognition (ASR) system or Speech Enhancement system. To enhance the performance of the speech processing system. In the literature related to azimuth detection of a source, the proposed technique can be roughly divided into two categories. One of them is to receive the data received by the microphone array, and then use beamforming or subspace theory to obtain the sound. The angle of the source. The other is to try to estimate the time delay (TD0A), and then use the relationship between the sound source and the wheat to estimate the angle of the sound source.戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 戍 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声 声Firstly, a voice signal from the sound 201101852 is received via a microphone array, wherein the microphone is received by each microphone to include a plurality of microphones. Then, wave. Next, each filtered signal is amplified and apostrophed. Then, by calculating each (four)^input (4) into a digital-to-digital ratio: the expression ===, the timbre of the closing function of the phonogram is used to calculate the estimated value of two = and form a TDOA vector: Ο 〇 TD 〇 A vector to perform azimuth calculation, with 4 - azimuth of the above-mentioned sound source and - elevation angle. (4) The invention provides a sound source detecting device for detecting the orientation of a sound source. The sound source preparation device comprises a microphone array including a first microphone, a second microphone, and a microphone for receiving a voice from the sound source: the second microphone = the second input The signal and the τ third input signal corresponding to the above-mentioned first flute-i. The above judging circuit firstly outputs the voice frame of the upper voice (the voice box containing the voice), and the above frame: when: the first: the voice box of the voice and the voice of the third input signal are first and The third input signal simultaneously detects the speech two-section function, and the judging circuit performs the estimation of the time delay TD0A value between the first, second, and third rounds, and then calculates one of the top meters. The above-mentioned and other objects, features, and advantages of the present invention will become more apparent. The following detailed description of the preferred embodiments and the accompanying drawings The sound source detecting device 100 is used to detect the orientation of the sound source 140. The sound source detecting device 100 includes the microphone 1, the array 110. The first embodiment shows the sound source detecting device 100 according to an embodiment of the invention. The signal conversion circuit 12A and the determination circuit 130, the signal conversion circuit m includes an amplification and chopping unit 122 and an analog-to-digital conversion unit 124, and the determination circuit 13 includes a voice activity price measurement time 70 132, a time delay estimation unit. 134 and square The calculating unit 136. The sound source detecting device 100 can obtain the three-dimensional orientation of the sound source 140 according to the voice signal Ss emitted by the sound source 140, that is, the horizontal azimuth angle of the sound source 140 and the elevation angle. Fig. 2 is a schematic view showing the structure of the microphone array 110 in Fig. 1. The microphone array 110 is composed of three microphones 210, 220 and 230, wherein three microphones 210, 220 and 230 are arranged in an equilateral triangle to form a In addition, the microphones 210, 220, and 230 are all omnidirectional microphones. Referring to Figures 1 and 2, when the voice signal Ss is transmitted from the sound source 14 to the sound source detecting device 1 0, the microphones 210, 22〇 and 23〇 respectively generate the corresponding input signals. For example, when the voice signal s is received, the microphone 210 provides the input signal Sin1 to the signal conversion circuit: the microphone 220 provides the input signal Sw to The signal conversion circuit 12A, and the microphone 230 provides the input signal Sin3 to the signal conversion circuit 12A. Since the distance between the sound source 140 and the microphones 210, 220 and 230 is not the same ', therefore, when the voice signal When the Ss is transmitted from the sound source 140 to the sound source detecting device 1 , the voice signal Ss does not reach the microphones 210, 220, and 23 同时 at the same time. For example, the distance between the microphone 210 and the sound source 140 is the closest to the microphone 23 〇201101852 The farthest distance from the sound source 140, the microphone 21〇 will receive the voice signal Ss first and the microphone 230 will finally receive the voice signal ^. ❹ ❹ Figure 3 shows the signal conversion circuit in Figure 1 A schematic diagram of 120. As shown in FIG. 3, amplification and filtering unit 122 includes amplifiers 302, 304, and 306 and filters 312, 314, and 316, and analog-to-digital conversion unit 124 includes three analog-to-digital converters 322, 324. And 326, wherein filters 312, 314, and 316 are low pass filters. Referring to the first and third figures, the input signals Sin1, Sw and Sho provided by the microphones in the microphone array 110 are respectively amplified and filtered by the corresponding amplifiers and subtractors. For example, the input signal Sini will be amplified by the amplifier 302 first, and then filtered by the filter 312 to obtain the filtered signal sfl. Similarly, the input signal Sin; 2 is amplified and filtered by the amplifier 3 〇 4 and the filter 314 to obtain a filtered signal, and the input signal, Sin3 is amplified and amplified by the amplifier 3 〇 6 and the filter 316 to obtain a filter. Signal SD. Then, the analog-to-digital converter %] Π will convert the filter money ^^ and ^ into a number

D1, D2, and D3 are transmitted to the judging circuit 13A. I refer back to Fig. 1, in the voice activity detecting unit 132 voice activity detection (v〇ice activity -, Vad: the signal frames of the bit signals Di, D2, and D3 are numbers. The voice activity side is a voice) Signal processing method == Whether the sound box contains spoken speech. By judging ===3 one-step processing, such as speech recognition, speech 'I' band: the rate is lower, that is, the energy of the speech signal is determined by the round The incoming signal is a voice or non-speech signal. First cloth; = motion: 7 201101852

The measuring unit 132 divides the digital signals D1, D2, and D3 into a plurality of frames. For example, assuming that the length of one frame is 23 ms and the length of the digital signal D1 has been received is 230 ms, the digital signal D1 can be divided into 10 frames. Then, each frame is first subjected to Fast Fourier Transform (FFT) to be converted to the frequency domain, and then the probability value of each band is defined, and then the spectral entropy of the frame is calculated. The low frequency component of the speech signal is strong (about concentrated below 3 kHz) and the intensity on the spectrum will show significant fluctuations. However, the spectral intensity of the noise signal changes less and is flatter. Therefore, by comparing the spectral disorder of each frame with a specific ambiguity reference value, the voice activity detecting unit 132 can determine whether the sound box is a voice box or a non-speech box. Then, when the sound box is determined to be a voice sound box, the voice activity detecting unit 132 further calculates a signal to noise ratio SNR of the sound box to avoid false positives, so as to improve voice activity detection and judgment. The accuracy. Figure 4 is a schematic diagram showing the estimation of the value of the Time Delay of Arrivd (TDOA) of two microphones using the principle of a generalized cross correlation function according to an embodiment of the present invention. We will make this theory even more ambiguous: an approaching approach. In Figure 4, the signals & and & microphones receive and amplify, chopped signals. For example, it is the output signal of the amplification and filtering unit in FIG. 3, the first signal, the 2 and the output signal s. Blind without 'L *5 Tiger X〗 and Hex σϊ, 4?n 2 θ knife passes through the linear time-invariant filter 410 and two, Λ ^yi and ^ Then, the multiplier _ multiplies the signal 仃A signal M is obtained in which the money yj signal 201101852 == is generated by the 430. Then, within the inter-turn translation range of the integral, the signal is “peak detector 460 ^ 5 flat one win — political μ target to find the maximum value of the generalized cross-correlation function. Two gates: =:: use the fourth graph The time delay estimation value 6 obtained by the generalized cross-correlation will be, for example, the power frequency r mi3. . , the 乂 cross-correlation function Κχι, Χ 2 (τ) and the interaction force drought frequency. The day Gxl, x2 (f) The relationship is: , (7) = ~ r (1) 〇Γ艮 = 4 graph can get the interactive power spectrum between the letter ", y2"

The relationship between Gyi, y2(f) and Gxi, x2(f): , G-(7), /)坧(/), (7).

= indicates a complex number of vehicles, and the miscellaneous earning is the frequency response of the filter 241〇 and Tao :=. Thus, the signal ~ and J correlation functions can be defined as the following equation (3) ·· m 〇)=]>"/)~(/) eJ^KfT ^ (7)'剌/). Because of the limited signal ~ and ( 3 heart) observation 〇 2 / should only be the wait, (/) estimated value "(7). Therefore, the formula U (3) can be rewritten into the following formula (4) · · fly t (Γ) = · 0" /) "J/V~y in the case of meaning (7)' (7). In addition, there is a large difference in the value of the value U of the φ 第 第 适当 适当 适当 适当 适当 适当 。 。 。 。 。. The frequency response of V: f can be made directly at phIt 11^ —Γ-, and the inter-correlation function between equations (5) below forms a pulse function, and the maximum value of the pulse function appears in the ruthless delay position. 9 201101852 Ψκ (/): 'ΚΓωί (5). Next, substituting equation (5) into I-'h two public) can obtain the following formula (6): 1 ^ 'When voice activity (4) unit 132 detects digital states t唬Dl, D2, and 乜3 The digital signal frame of the aj digit Ο 进行 performs a = conversion for each microphone wheel to obtain the time delay of each frequency in the frequency domain. In fact, the right and left winds 210, 220 * 230 are equal to the left and right _ Shishi is 'first according to the formula (6) and D (ie, the 〇, the ▲ ▲ do not get Dl (ie VO digital signal) the sound box is not (,) 6(7) and Ιά r in the fast Fourier transform spectrum of the sound box of "(:): followed by bit 2), and the spectrum is not (/) and especially 2 (/) to approximate the equation (6) g(7), that is, by {(7)"(7) approximation (6) Ι> Λ Λ / / "14^).5(/)1 is approximated by Κ (/)|; then, the inverse of the discrete value is counted The fast Fourier transform operation evaluates to 'find the maximum: the value; then, the reasonable range value from the Γ value. Estimate a value at time delay, as the estimated time-interval correlation function of the time delay:: 134' It is by using the above-mentioned widely used method, the value of the time delay is estimated every time. The delay is estimated. The input signal of the wind can get a time, and the total is available, for the second In the figure, the microphones 210, 220 and 230 have a time delay estimate (four) even, which can form a reasonable range of time delay values between the two microphones. The symptom is determined by the transmission speed of the sound in the air in 201101852. Therefore, when the position of the sound source is in line with the two microphones, the maximum time delay estimate can be obtained. In addition, when the (from () is obtained The larger the value of time delay 7 is, the larger the difference between the angle values corresponding to the two adjacent discrete values of τ is. Therefore, in order to improve the accuracy of the time delay estimation, we use parabolic interpolation. (parabola interpolation) to obtain the τ value corresponding to the parabola vertex to reduce the angular error. " Ο 〇 Figure 5 shows the position map between a sound source 530 and a microphone pair (ie, Mike and 520). (4) 53〇 The three-dimensional coordinates, the winds 51. and 52. The three-dimensional coordinates are {~, from the second} and =, according to the following formula (7) and formula (8), the sound source can be transmitted by tl w (7) (8). =Κι - λ)2+k, -zy t2 = ysf + c \〇) ί The speed at which the two sound waves propagate in the air. Therefore, the time delay between the microphone 510 and the microphone 520 can be calculated for a time. Delaying the recognition of the mother microphone Value vector (A A. will extend the theoretical value of the formula - the formula (1), the azimuth calculation map, calculate - a pair::: = alignment, straight (four) flat angle: and elevation angle 兀 136 - total Obtaining [inverse y ^ and thus the orientation +1 theoretical value vector. Next, the orientation 201101852 calculus 7G 136 calculates the time delay estimation vector (M4) from the time delay estimation unit 134 and all the time-late theoretical vector vectors. The geometric distance 'to find the theoretical vector of the time delay with the smallest distance, and then the azimuth angle and elevation angle 声 of the sound source can be obtained. As described, the sound source pick-up device described in the present invention can use an array of equilateral triangles composed of microphones to detect the three-dimensional orientation of the sound source. In the present invention, speech activity detection based on chaotic measurement and time delay estimation based on the generalized interactive correlation function principle are used, thereby reducing the horizontal azimuth of the sound source and the error in estimation of the elevation angle. The present invention has been described above with reference to the preferred embodiments. However, it is not intended to limit the scope of the invention, and it is possible to make some modifications and refinements without departing from the scope of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims.

12·201101852 [Simplified description of the drawings] Fig. 1 shows a sound source detecting device for detecting the orientation of a sound source according to an embodiment of the present invention; and Fig. 2 shows a microphone array of Fig. 1 FIG. 3 is a schematic diagram showing a signal conversion circuit in FIG. 1; FIG. 4 is a schematic diagram showing estimating a time delay value between two microphones using a generalized cross-correlation function principle according to an embodiment of the invention; And Figure 5 shows a positional map between a sound source and a pair of microphones. 〇 [Main component symbol description] 100~ sound source detection device; 110~ microphone array; 120~ signal conversion circuit; 122~ amplification and filtering unit; 124~ analog-to-digital conversion unit; 130~ judgment circuit; Detecting unit; 134~time delay estimating unit; 136~azimuth calculating unit; 140, 530~ sound source; 210, 220, 230, 510, 520~ microphone; 302, 304, 306~ amplifier; 312, 314, 316~ Filter; 〇322, 324, 326~ analog-to-digital converter; 410, 420~ linear time-invariant filter; 430~ retarder; 440~multiplier; 450~ integrator; 460~peak detector;

Dl, D2, D3 ~ digital signal; d, A, 4, A ~ time delay estimation value;

Sfl, Sf2, Sf3 ~ filter, wave signal;

Sinl, Si„2, Sjn3, Χι, X2~ input signals;

Ss ~ speech signal; y], y2, y3, Μ ~ signal; 61 ~ azimuth; and more ~ elevation angle. 13

Claims (1)

201101852 VII. The scope of application for patents······································································································· , including: 1.- An ancient detection method, suitable for finding a microphone array, picking up your J, the sound source, the orientation, each of the above microphones: the wind; the mouth filtering, and the magnified The signal is amplified and the signal is converted into a signal; the input signal is converted into a digital signal Ο 音 Ξ ί ί ί ί ί ί ί ί ί ί ί ί ί ί ί : ί , , 语音 = = = = = = = = = = = = = = = = = = = = = = = = = Azimuth and azimuth calculation to obtain the above-mentioned complex digital signal m sound source detecting method 'where the speech signal is detected by the spectral disorder of the speech sound box of the letter "two = sound = bracket:" The second item of the above plurality of digit signals is used for voice activity, wherein the pair of == complex sounds (four) of the signal-to-synchronization and the voice signal is calculated by the above-mentioned time measurement method, wherein the above-mentioned conversion is determined as each value is obtained; The generalized cross-correlation function between 〃 and 1414 201101852 estimates the vector of the above-mentioned time delay value according to the generalized cross-correlation function between the above complex microphones. 5. The sound source detection as described in claim 1 The method, wherein the performing the azimuth calculation step further comprises: calculating a difference between the time delay estimated value vector and a specific vector to obtain the azimuth angle and the elevation angle. 6. The sound source detection as described in claim 1 The method of measuring, wherein the microphone array comprises a first microphone, a second microphone and a third microphone, wherein the first, second and third microphones are arranged in a planar array of a ^ equilateral triangles. The sound source detecting method of the sixth aspect of the invention, wherein the first, second, and third microphones are omnidirectional microphones. 8. A sound source detecting device, configured to detect a sound source orientation The method includes: a microphone array including a first microphone, a second microphone, and a third microphone for receiving the sound from the sound a voice signal to obtain a first input signal corresponding to one of the first microphones, a second input signal corresponding to one of the second microphones, and a third input signal corresponding to one of the third microphones; and a determination a circuit for detecting a voice frame of the first input signal, a voice frame of the second input signal, and a voice frame of the third input signal, and simultaneously detecting the first and second When the voice box of the third input signal is used, calculating a correlation function value between the first, second, and third input signals to obtain a horizontal azimuth and an elevation angle of the sound source. The sound source detecting device of item 8, wherein the microphone array is a planar array in which the first, second, and third microphones are arranged as an equilateral triangle of 15 - 201101852. 10. The sound source detecting device of claim 9, wherein the first, second, and third microphones are omnidirectional microphones. 11. The sound source detecting device of claim 8, further comprising: a signal conversion circuit coupled between the microphone array and the determining circuit, comprising: an amplifying and filtering unit for amplifying The first, second, and third input signals, and the low-pass filtering of the amplified first, second, and third input signals; and a type of comparison digital conversion unit for filtering the above The first, second and third input signals are respectively converted into a first digital signal, a second digital signal and a third digital signal. 12. The sound source detecting device of claim 11, wherein the determining circuit comprises: a voice activity detecting unit, configured to divide the first, second, and third digit signals into a plurality of The sound box, and detecting the voice frames in the first, second and third input signals respectively according to the amount of spectrum disturbance. 13. The sound source detecting device according to claim 12, wherein the voice activity detecting unit further sets a voice frame in the first, second, and third input signals according to a signal noise ratio. Further confirmation as a voice box. 14. The sound source detecting device of claim 12, wherein the determining circuit further comprises: a time delay estimating unit, wherein when the voice activity is detected, the detecting unit 16 201101852 detects the first When the sound box in the second and third input signals is a voice box, the time delay estimating unit performs fast Fourier transform on the first, second and third digit signal frames, and uses the Fourier spectrum of each input signal respectively. Obtaining an interaction power spectrum between the two signals to obtain an approximation value of the correlation function between the two microphones of the first, second, and third microphones, and estimating the voice according to the approximation value of the correlation function The signal arrives at a time delay value vector between the two microphones of the first, second and third microphones. 15. The sound source detecting device of claim 14, wherein the correlation function is a generalized cross-correlation function between the two input signals. 16. The sound source detecting device of claim 14, wherein the determining circuit further comprises: an orientation calculating unit configured to obtain the horizontal azimuth angle according to the time delay estimated value vector and a specific vector, and The above elevation angle. ❹ 17