TW201303854A - A method for noise reduction and speech enhancement - Google Patents

Abstract

A method for noise reduction and speech enhancement includes steps of providing at least two microphones to receive at least two microphone signals, transferring the two microphone signals into frequency domain by fast Fourier transform (FFT) to obtain a speech signal and a noise signal therein, calculating an angle between the speech and noise signal, and further obtaining an ITD by applying a phase difference estimation algorithm to the angle, calculating a threshold value of the ITD corresponding to the angle, applying a binary mask principle to the ITD and its threshold value to purify the speech signal out of the noise signal, and transferring the speech signal back into time domain to obtain a purified speech signal, thereby removing the noise signal and enhancing the speech quality.

Description

Method for eliminating noise and improving speech recognition rate

The invention relates to a method for eliminating microphone noise, in particular to a method for eliminating noise and reducing echo interference in communication, so as to effectively improve the speech recognition rate.

Generally speaking, the way in which the microphone receives the sound signal is mainly divided into single channel and dual channel. Among them, the single channel denoising method needs to estimate the noise cancellation ratio, and the dual channel sensing mostly uses beam forming. A directional microphone system is produced in an array.

Such a microphone system is highly sensitive to human voice, and thus receives a sound signal from a person's position and is less sensitive to background noise. However, since such a microphone system includes two or more microphones, the beam formed by the microphone system is quite large, which easily causes a problem of insufficient directivity.

Mobile phone communication noise cancellation devices currently used in vehicles or in general indoor use a large number of microphones, various filters and large matrix operations. With such a heavy amount of computing, huge memory space and numerous microphones, the cost of hardware is a big burden.

Secondly, due to the lack of directivity, neither the products on the market nor the patents and literature on microphone arrays can effectively eliminate noise and not distort the speech in the presence of noise.

In addition, general mobile phones or in-vehicle communication devices often have problems in that the echo is too large during the call, which affects the communication quality.

Therefore, how to propose a microphone receiving method that can effectively eliminate noise in the environment and improve voice quality is one of the problems that need to be solved by those skilled in the art.

The main object of the present invention is to provide a method for eliminating noise and improving speech recognition rate, which uses a golden ratio search method with Taylor's theory to calculate an optimal interaural time difference threshold, so that the voice signals of each angle are Get the best voice quality.

Another object of the present invention is to provide a method for eliminating noise and improving speech recognition rate by using a composite echo cancellation system to filter out the main acoustic echo and environmental disturbance of the voice signal, thereby eliminating the communication process. The resulting echo further enhances voice quality.

In order to achieve the above object, the present invention relates to a method for eliminating noise and improving speech recognition rate, comprising the steps of: providing two or more microphones for receiving at least two microphone signals; and utilizing the microphone signals quickly Fourier transforms to the frequency domain to obtain one of the voice signals and a noise signal; calculates the angle between the voice signal and the noise signal, and uses a phase difference algorithm to further find the time difference between the ears; according to the voice signal and the noise signal The angle is calculated as a threshold value of the time difference between the ears; according to the time difference between the ear and the threshold, a masking rule is used to obtain the voice signal to remove the noise signal; and the voice signal is transferred to the time by using an inverse fast Fourier transform and superposition module. Domain output.

The invention further relates to a method for eliminating noise and improving speech recognition rate, comprising the steps of: providing two or more microphones for receiving at least two microphone signals; and converting the microphone signals into a frequency domain by using fast Fourier transform, Obtaining a voice signal and a noise signal in the microphone signal; calculating an angle between the voice signal and the noise signal, and using a phase difference algorithm according to the angle to match the shadow estimation to obtain the voice signal in the microphone signal, and removing the noise signal The voice signal is converted to the time domain output by using an inverse fast Fourier transform and superposition module; and a composite echo cancellation system is coupled to the voice signal of the time domain to filter out the acoustic disturbance of the voice signal.

The purpose, technical contents, features and effects achieved by the present invention will be more readily understood by the detailed description of the embodiments and the accompanying drawings.

The invention provides a method for eliminating noise and improving speech recognition rate, which utilizes a phase difference between two microphones to obtain a mask of a microphone signal in a time domain and a frequency domain, and eliminates noise to improve speech quality.

Please refer to FIG. 1 , which is a schematic diagram of a microphone array capable of eliminating noise and improving speech recognition rate according to an embodiment of the present invention, including a microphone array (including at least two microphones 14 , 14 ′′) and at least two fast Fourier transform modules. 16, 16', a computing module 18, a mask estimation module 20, and an inverse fast Fourier transform and overlay module 22.

Please refer to FIG. 2, which is a flow chart showing the steps of a method for eliminating noise and improving speech recognition rate according to an embodiment of the present invention. For the description of the embodiments of this embodiment, please refer to the figures 1 to 2 together.

After the sounds of the voice source 10 and the noise source 12 are transmitted as shown in step S202, the microphones 14, 14' receive the microphone signals containing both the noise signals and the voice signals.

Then, as shown in step S204, the fast Fourier transform modules 16, 16' are used to convert the microphone signals received by the microphones 14, 14' to the frequency domain to obtain the voice signals and noise signals in the microphone signals.

Then, as shown in step S206, the computing module 18 is connected to the microphones 14, 14' for calculating the angle between the voice signal and the noise signal in the microphone signal. Thereby, the computing module 18 further uses the phase difference algorithm to find the interaural time difference (ITD) according to the angle.

As shown in step S208, after the computing module 18 finds the time difference between the ears, the computing module 18 further calculates the threshold of the angle between the inter-aural time corresponding to each of the noise signals and the voice signal.

Then, as shown in step S210, the mask estimation module 20 uses a masking rule to obtain a voice signal and remove the noise signal according to the calculated time difference between the ear and the threshold.

Finally, as shown in step S212, the inverse fast Fourier transform and superposition module 22 is configured to convert the voice signal from the frequency domain back to the time domain to obtain a voice signal having a higher speech recognition rate after removing the noise.

In step S204, after the noise signal and the voice signal are received via the microphones 14, 14', the fast Fourier transform module 16, 16' is transferred to the frequency domain via a Hamming window and a fast Fourier transform (FFT). The second microphone signals P ₁ ( k,l ) and P ₂ ( k,l ) are as shown in the following equations (1) and (2):

Wherein (k, l) represents the k th frequency, the l th frame, X represents a voice signal, N _i represents the i-th noise sources, the received signal P _m is the m-th microphone, ω _k = 2πk / N, 0 ≦ k ≦ N / 2-1, N is the length of the fast Fourier transform.

Next, in step S206, the computing module 18 calculates an angle between the noise signal and the voice signal in the two microphone signals P ₁ ( k, l ) and P ₂ ( k, l ), that is, between the voice source 10 and the noise source 12 . The angle of the angle to further find the time difference between the ears (ITD).

Generally speaking, if the voice signal is directly in front of the microphone, the time difference between the ears is 0. The noise in other directions uses d _i ( k, l ) to indicate the time difference between the ears. The time difference between the ears is related to time and frequency. . If the time-frequency domain bin( k _j , l _j ) is dominated by a strongest interference, the above equations (1) and (2) can be simplified to the following equations (3) and (4):

The time difference between the ears at this time can be obtained by calculating the phase difference between the two microphone signals, as shown in the following equation (5):

Then, in step S208, the computing module 18 further calculates the threshold of the angle between the inter-aural time difference corresponding to the noise signal and the voice signal. According to an embodiment of the present invention, the calculation module 18 calculates the optimal threshold value by using the Golden-Section Search (GSS) with Taylor's theory to find the optimal threshold τ corresponding to each angle.

Suppose a function f(x) is continuous and has a minimum value in [a, b], and two points c and d are selected in [a, b], and the relationship is as follows:

Where d is c in The symmetry point on the line segment compares the size of f(c) and f(d). If f(c)<f(d), the new search point becomes [a,d], otherwise it becomes [c,b], Then take another point in the new range, compare the size of the two internal points again, repeat this step to continue to narrow the range, when the range is small enough to accept the point, treat it as a function f(x) at [a, b] the minimum value of the interval, according to Taylor's theory, when the function f(x) is close to x _m , its value approximates:

If f(x) is close enough to f(x _m ), the subsequent second derivative term is negligibly small, so equation (10) can be expressed as the following equation (11):

Where ε is 10 ^-3 . Using the speech distortion, denoising degree and overall speech quality as the parameters of the function in the golden ratio search method, the function of the angle τ value can be obtained as follows (12):

τ(i)=(-7.76*10 ^-5 )i ² +(1.69*10 ^-2 )i-(5.45*10 ^-2 ) (12)

Where i is the angle between the voice signal and the noise signal, and the threshold τ corresponding to the angle i can make the processed signal have the best voice quality.

Therefore, after obtaining the optimal threshold value τ of the time difference between the ears, in step S210, the mask estimation module 20 estimates the masking signal of the microphone signal according to the binary mask principle by the following formula (6):

Among them, only the signal whose time difference between the ears is smaller than τ will be regarded as the target voice signal.

The last voice signal S( k , l ) can be averaged by the two microphone signals. ( k , l ) and the masking signal B( kj,lj ) are multiplied, as shown in the following equation (7) and (8):

After the voice signal is obtained in step S210 to be successfully separated from the noise signal, in step S212, the inverse fast Fourier transform and superposition module 22 performs the inverse fast Fourier transform (IFFT) and the overlap addition method on the voice signal in the frequency domain. (OLA) is converted to time domain signal output to obtain a voice signal with a higher speech recognition rate after noise removal.

Please refer to FIG. 3, which is a schematic diagram of a microphone array capable of eliminating noise and improving speech recognition rate according to another embodiment of the present invention. As shown in FIG. 3, in the framework of the present invention, the inverse fast Fourier transform and superposition module 22 can be further connected with an automatic speech recognition module 24 for receiving the output of the inverse fast Fourier transform and superposition module 22. Voice signal for speech recognition.

Secondly, considering that if the sound source position is not directly in front of the microphone array, the present invention further proposes a beam-steering technique for controlling the beam steering of the microphone by adding different delays to the respective microphones. Angle to turn it to the sound source position.

Assuming that the steering angle is θ _M , the filter frequency factor of the beam steering can be as shown in the following equation (13):

Where n is the nth microphone, ω is the frequency factor, f _s is the sampling frequency, and d is the microphone spacing. Then in the time domain, the filter can be written as a delay according to the following equation (14):

Since the delay of the above formula is not an integer, Lagrange interpolation must be used to make it easier to achieve. This interpolation method can use the Infinite Impulse Response Filter shown in the following formula (15). Simple completion:

Where N is the order of the filter, where the first order is used, and D is the delay fraction.

According to an embodiment of the invention, the angle of beam steering includes 0 degrees to 180 degrees. That is to say, after the microphone array receives the microphone signal, the microphone first performs omnidirectional (0°~180°) beam steering, and after each beam steering, performs spectrum analysis to calculate the time difference between the ears, and then passes the above formula ( 6) The masking rule preserves the target sound source and suppresses interference. After the above process of voice purification, the beam energy of each microphone at each steering angle is finally calculated for the direction of arrival estimation (DOA) of the voice source.

The reason is that when the microphone is turned to the actual sound source orientation, the maximum energy should be available (because the energy of the target sound source can pass the masking rule of the above formula (6)), thereby judging the correct sound source direction. . It can calculate its energy level by the following formula (16):

among them ( k , l ) is the signal after the two-channel signal is purified by the phase difference algorithm; ( k , l ) is the frequency and time index respectively; e ^-jkλ is the beam steering filter as a function of frequency; and λ must be As shown in the following equation (17), between the maximum and minimum delay times:

Please refer to FIG. 4, which is a schematic diagram of a microphone array capable of eliminating noise and improving speech recognition rate according to another embodiment of the present invention, including a microphone array (including at least two microphones 14, 14') and at least two fast Fourier transforms. The modules 16, 16', the computing module 18, the shadow estimation module 20, the inverse fast Fourier transform and superposition module 22, a stationary filter 26 and an adaptive filter 28.

Please refer to FIG. 5, which is a flow chart showing the steps of a method for eliminating noise and improving speech recognition rate according to another embodiment of the present invention. For the description of this embodiment, please refer to the figures 4 to 5 together.

Steps S502 to S508 are the same as steps S202 to S212 of the previous embodiment of the present invention, and therefore will not be repeated herein. It should be noted that, in this embodiment, the present invention further includes step S510: connecting a composite echo cancellation system (the fixed filter 26 and the adaptive filter 28) to convert the voice signal back to the time domain to utilize This composite echo cancellation system filters out acoustic disturbances of the voice signal.

In detail, when the user talks with a remote third party and the system has a speaker 30, and the remote third party generates a far-end audio 32, the speaker 30 and the microphone 14, 14' will form a Fixed echo path. The present invention is followed by a fixed filter 26 and an adaptive filter 28 in an inverse fast Fourier transform and superposition module 22 to produce a composite echo cancellation system.

In this embodiment, the fixed filter 26 is used to filter out the main acoustic echo of the voice signal, and the adaptive filter 28 is used to filter out the problem caused by the disturbance of the voice signal in the surrounding environment. For example, a fixed filter can first identify the dynamic characteristics of the system offline (off-line) before joining the system. The adaptive algorithm of the composite system may be, but not limited to, a filterd-x LMS algorithm, which includes: an entire echo cancellation path as shown in the following formula (18):

f ( n ) is a fixed filter, w ( n ) is an adaptive filter, where w ₀ ( n ) is 1, Δ w ( n )=[ w ₁ ( n ) w ₂ ( n )... w _{L - 1} ( n )], δ( n ) is a unit pulse sequence; the calculation error signal as shown in the following equation (19):

e ( n )= d ( n )- y ( n )= d ( n )- w ^T ( n )[ f ( n )* x ( n )] (19)

d ( n ) is the microphone income signal, y ( n ) is the filter output signal, w ( n )=[ w ₁ ( n ) w ₂ ( n )... w _{L -1} ( n )] ^T is at time n Adaptive filter coefficient composition vector, x ( n )=[ x ( n ) x ( n -1)... x ( n - L +1)] ^T is the input signal vector at time n ; and according to the following formula (20) Using the FXLMS algorithm to update the adaptive filter,

However, it is worth noting that when the far-end third party generates the far-end audio 32, the signals received by the microphones 14, 14' at this time will not only be the acoustic echo generated by the system, but will cause an adaptive filter. 28 divergence. In view of this, as shown in FIG. 6, the present invention further includes steps S602 to S608 before determining to update the adaptive filter 28.

As shown in step S602 to step S604, the system includes a double talk detector (DTD) for detecting whether the voice signal generated by the voice source 10 and the far-end audio 32 generated by the remote third party are At the same time. Thereafter, as shown in step S606, if both occur simultaneously (that is, the user simultaneously speaks with the remote third party), the adaptive filter 28 is stopped. Otherwise, as shown in step S608, if the two do not occur simultaneously (that is, the user does not simultaneously speak with the remote third party), the adaptive filter 28 continues to be continuously updated.

In particular, the present invention primarily compares the signals received by the microphones 14, 14' with the signals output by the adaptive filter 28. Since the comparison energy amount causes the adaptive filter 28 to switch too much, the microphone signal d ( n ), the fixed filter output signal x '( n ) and the fixed filter are calculated according to the following equations (21) to (23). The adaptive filter outputs a signal y ( n ) to form a wave seal v _d ( n ), v _x ( n ) and v _y ( n ), α = 0.99.

v _x ( n )= αv _x ( n -1)+(1 - α )| x ( n )| (21)

v _d ( n )= αv _d ( n -1)+(1 - α )| d ( n )| (22)

v _y ( n )= αv _y ( n -1)+(1 - α )| y ( n )| (23)

According to the following formulas (24) and (25), the detection function ξ( n ) and the dynamic threshold function are obtained from v _d ( n ), v _x ( n ) and v _y ( n ). Function) T ( n ). When the detection function ξ( n ) is greater than the dynamic threshold function T ( n ), the remote third party generates the far-end audio 32, and the adaptive filter 28 also stops updating.

When the detection function ξ( n ) is less than the dynamic threshold function T ( n ), the adaptive filter 28 continues to update. The optimum value of γ is 0.05 in the experiment, and the addition of the small positive real number β is the range reserved for preventing detection errors.

In summary, the present invention provides a method for eliminating noise and improving speech recognition rate, which can implement an acoustic signal processing method in a telecommunication communication system. In this way, not only the phase difference between the two microphones can be utilized, but also the sound source angle is obtained to determine the beam opening size, so as to improve the speech recognition rate, and the sound source position can be automatically detected through the beam steering. In addition, beam steering technology can be used to solve the problem that the voice signal is not at the spindle position.

The invention provides a method for eliminating noise and improving the speech recognition rate, and can be applied to a barge in system, and combined with a composite echo cancellation system, effectively reducing the interference of the echo on the recognition rate. This system is suitable for mobile phones, smart toys and other instruments that need to be used in the speech recognition system, so that the identification system can have a good recognition rate even in a space with loud noise and reverberation.

The embodiments described above are merely illustrative of the technical spirit and the features of the present invention, and the objects of the present invention can be understood by those skilled in the art, and the scope of the present invention cannot be limited thereto. That is, the equivalent variations or modifications made by the spirit of the present invention should still be included in the scope of the present invention.

10. . . Voice source

12. . . Noise source

14, 14’. . . microphone

16, 16’. . . Fast Fourier Transform Module

18. . . Computing module

20. . . Mask estimation module

twenty two. . . Anti-fast Fourier transform and overlay module

twenty four. . . Automatic speech recognition module

26. . . Fixed filter

28. . . Adaptive filter

30. . . speaker

32. . . Remote audio

1 is a schematic diagram of a microphone array capable of eliminating noise and improving speech recognition rate according to an embodiment of the present invention.

2 is a flow chart showing the steps of a method for eliminating noise and improving speech recognition rate according to an embodiment of the present invention.

3 is a schematic diagram of a microphone array capable of eliminating noise and improving speech recognition rate according to another embodiment of the present invention.

4 is a schematic diagram of a microphone array capable of eliminating noise and improving speech recognition rate according to another embodiment of the present invention.

Figure 5 is a flow chart showing the steps of a method for eliminating noise and improving speech recognition rate according to another embodiment of the present invention.

Figure 6 is a flow chart showing the steps before updating the adaptive filter according to Figure 5.

Claims (20)

A method for eliminating noise and improving speech recognition rate, comprising the steps of: providing two or more microphones for receiving at least two microphone signals; converting the microphone signals to a frequency domain by using fast Fourier to obtain the microphone signals a voice signal and a noise signal; calculating an angle between the voice signal and the noise signal, and using a phase difference algorithm to further find an inter-ear time difference; calculating the angle according to the voice signal and the noise signal a threshold value of the time difference between the ears; according to the time difference between the ear and the threshold, a masking rule is used to obtain the voice signal to remove the noise signal; and the voice signal is subjected to an inverse fast Fourier transform and superposition mode The group goes to the time domain output. A method for eliminating noise and improving a speech recognition rate as described in claim 1, wherein the calculation of the threshold is performed by using a Golden-Section Search with Taylor's theory. The method of claim 2, which can eliminate noise and improve speech recognition rate, wherein the golden ratio search method selects two points in a continuous range, and compares the function value of one of the two points to reduce the continuous range. And repeating the steps of optionally selecting two points and comparing the function values to continue narrowing the continuous range to find a minimum value of the function value in the continuous range, the threshold value being obtained by using the minimum value in conjunction with Taylor's theory. The method of claim 1, wherein the masking rule further comprises the steps of: comparing the time difference between the ear and the threshold to obtain a masking signal; and The average of the microphone signals is multiplied by the masking signal to obtain the voice signal in the microphone signals. The method of claim 1, wherein the inverse fast Fourier transform and superposition module converts the voice signal in the frequency domain into a time domain by using an inverse fast Fourier transform and an overlap addition method. Signal. The method of claim 1, which can eliminate noise and improve speech recognition rate, wherein when the voice signal is located directly in front of the microphones, the time difference between the ears is zero. The method of claim 1, wherein the microphone signal is regarded as the voice signal when the time difference between the ears is less than the threshold. The method of claim 1, wherein the microphones are arranged in an array. The method for eliminating noise and improving the speech recognition rate according to claim 1, further comprising receiving the voice signal output by the inverse fast Fourier transform and superimposing module by using an automatic speech recognition module to perform speech recognition. The method of claim 1, wherein the receiving the microphone signals further comprises the step of: adding a delay to each of the microphones to control beam steering angles of the microphones. The method of claim 10, which can eliminate noise and improve speech recognition rate, wherein after adding the delay to each of the microphones, the method further includes the steps of: calculating beam energy of each of the microphones at each steering angle to determine the voice. The direction of the sound source of the signal. The method of claim 11, wherein the beam steering angle of the microphones is from 0 degrees to 180 degrees. The method of claim 1, wherein the voice signal is cancelled and the voice recognition rate is improved, wherein after converting the voice signal to the time domain, the method further comprises the step of: connecting a composite echo cancellation system to convert the voice signal back to the time domain. To filter out the acoustic disturbance of the voice signal. The method of claim 13, wherein the composite echo cancellation system comprises a fixed filter and an adaptive filter, wherein the fixed filter filters out the acoustics of the voice signal. Echo, the adaptive filter filters out the disturbance caused by the voice signal in the environment. The method of claim 14, which can eliminate noise and improve speech recognition rate, further includes the step of: updating the adaptive filter by using an FXLMS algorithm. The method of claim 15, wherein the method further comprises the steps of: providing a double talk detector (DTD); and detecting the noise filter. Whether the voice signal coincides with a far-end audio; and when the voice signal coincides with the far-end audio, the adaptation of the adaptive filter is stopped. A method for eliminating noise and improving speech recognition rate, comprising the steps of: providing two or more microphones for receiving at least two microphone signals; converting the microphone signals to a frequency domain by using fast Fourier to obtain the microphone signals a voice signal and a noise signal; calculating an angle between the voice signal and the noise signal, and using a phase difference algorithm according to the angle to match the shadow estimation to obtain the voice signal in the microphone signals, and removing The noise signal; the voice signal is transferred to the time domain output by using an inverse fast Fourier transform and superposition module; and the composite echo cancellation system is connected to the voice signal of the time domain to filter out the voice signal. Acoustic disturbance. The method of claim 17, wherein the composite echo cancellation system comprises a fixed filter and an adaptive filter, and the fixed filter filters out the acoustics of the voice signal. Echo, the adaptive filter filters out the disturbance caused by the voice signal in the environment. The method of claim 18, which can eliminate noise and improve speech recognition rate, further includes the step of: updating the adaptive filter by using an FXLMS algorithm. The method of claim 19, which can eliminate noise and improve speech recognition rate, wherein before updating the adaptive filter, the method further comprises the steps of: providing a double talk detector (DTD); detecting the Whether the voice signal coincides with a far-end audio; and when the voice signal coincides with the far-end audio, the adaptation of the adaptive filter is stopped.