An active noise control method based on sound source separation

Abstract

The present invention proposes an active noise control method based on sound source separation. When multiple sound sources emit sound simultaneously, the reference signal is first separated by using a time domain separation network, and then the main noise signal is input into the control system as a reference signal, which can achieve a greater noise reduction than the traditional method. The separation network can be used as a front-end module of the ANC system without occupying system resources.

Description

Technical Field

[0001] The invention relates to the field of active noise control and proposes an active noise control method based on sound source separation. Background Art

[0002] Noise pollution is a particularly prominent problem in environmental pollution. Noise control methods can be divided into passive noise control and active noise control (ANC). Passive noise control mainly relies on different sound insulation materials or sound-absorbing structures to isolate and absorb noise energy. The advantage of passive noise control is that it has a strong noise reduction effect on medium and high frequency noise, but the control performance of low-frequency noise is greatly reduced. For low-frequency noise, people proposed ANC, which adaptively realizes the control of primary noise based on the principle of destructive interference of sound waves.

[0003] A feedforward control system is commonly used in ANC systems, which requires a reference sensor to be placed near the noise source to obtain a reference signal. When there are multiple sound sources, the reference microphone picks up multiple noise signals at the same time, including the main noise signal and other interference signals. For example, the main noise signal is interfered by impulse noise or speech, which is not the main source of primary noise. In traditional methods, the reference signal is not processed, but directly used as the input of the system to participate in the iteration of the algorithm, which limits the amount of noise reduction. Summary of the invention

[0004] In view of the problems existing in the prior art, the present invention proposes an active noise control method based on sound source separation. When multiple sound sources emit sound simultaneously, the reference signal is first separated using a time domain separation network, and then the main noise signal is input into the control system as a reference signal, which can achieve a greater noise reduction than the traditional method. The separation network can be used as a front-end module of the ANC system without occupying system resources.

[0005] The technical solution of the present invention is:

[0006] An active noise control method based on sound source separation comprises the following steps:

[0007] Step 1: Build and train an end-to-end time-domain sound source separation system;

[0008] Step 2: Use the reference sensor to pick up the mixed reference signal x(n) emitted by multiple noise sources.

[0009] Step 3: Use the time domain sound source separation system established in step 1 to separate the mixed reference signal x(n) to obtain the signal x1(n) from the main noise source s1(n), which is used as the input of the control filter to obtain the secondary signal y(n).

[0010]

[0011] Where w(n) is the weight coefficient of the control filter;

[0012] Step 4: The secondary signal drives the secondary sound source to generate a cancellation signal y'(n), and the cancellation signal y'(n) and the sound pressure of the primary sound fields of the multiple noise sources at the error sensor are superimposed to form an error signal e(n);

[0013] Step 5: Iterate the control filter weights using the signal x1(n) of the primary noise source s1(n), the secondary path transfer function h(n), and the error signal e(n):

[0014] w(n+1)=w(n)-μr(n)e(n)

[0015] Where μ is the step size parameter, r(n)=x1(n)*h(n).

[0016] Furthermore, a deep learning-based network is used to construct the time domain sound source separation system; the time domain sound source separation system is constructed by an encoder-separator-decoder framework; wherein the encoder is implemented by a one-dimensional convolutional layer, the separator is used to estimate the masking matrix of the sound source, and the decoder is a one-dimensional deconvolutional layer; the masking matrix of each sound source from the separator is multiplied by the two-dimensional expression of the mixed sound from the encoder, and then the time domain signals of each sound source after separation are synthesized through the decoder.

[0017] Furthermore, the training process of the time domain sound source separation system is:

[0018] First, a training data set is established, wherein the training data set includes a plurality of noise sources; two different types of sound sources are randomly selected from the training data set, and mixed at a random signal-to-noise ratio to obtain a mixed sound source;

[0019] The obtained mixed sound source data is used to train the time domain separation network to obtain an end-to-end time domain sound source separation system.

[0020] Furthermore, during the training process of the time-domain sound source separation system, a scale-invariant signal-to-interference ratio is used as an evaluation index of the time-domain sound source separation system.

[0021] Beneficial Effects

[0022] The present invention has the following effects:

[0023] (1) The present invention uses a time domain separation network to separate the signal obtained by the reference sensor to obtain the main noise signal, which is then input into the control system as a reference signal. Compared with traditional methods, a greater amount of noise reduction can be achieved.

[0024] (2) The separation network can be implemented in the front-end module of the ANC system without occupying the resources of the control system.

[0025] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0027] Figure 1 Schematic diagram of ANC system based on sound source separation;

[0028] Figure 2 Separate network framework;

[0029] Figure 3 Weight coefficient iteration flow chart;

[0030] Figure 4 SI-SDR changes with the number of iterations on the training set and validation set;

[0031] Figure 5 Frequency response of the primary path; (a) P1(z), (b) P2(z);

[0032] Figure 6 Secondary path frequency response;

[0033] Figure 7 ANR for two noise reduction methods. DETAILED DESCRIPTION

[0034] The present invention mainly aims at the problem that when there are main noise sources and interference noise sources, the reference microphone of active noise control picks up multiple noise signals as reference signals at the same time, causing the main signal to be interfered by impulse noise or speech, and the actual noise reduction effect is poor. An active noise control method based on sound source separation is proposed.

[0035] ANC systems based on sound source separation such as Figure 1As shown. In the figure, x(n) is the reference sensor, which picks up the acoustic signals emitted by the noise sources s1(n) and s2(n) as the reference signal. Among them, s1(n) is the main noise source, and s2(n) is the interference noise source, that is, the amplitude of d1(n) is greater than d2(n). y(n) is the secondary signal calculated by the controller, which drives the secondary sound source after passing through the power amplifier, and generates a cancellation signal y'(n) at the error sensor. The primary sound field and the secondary sound field sound pressure are superimposed to form an error signal e(n). P1(z) and P2(z) are the primary paths of the two sound sources respectively, and H(z) represents the secondary path. w(n) represents the current filter weight coefficient value. Before the reference signal x(n) is input into the control filter, it first undergoes sound source separation to obtain the signal component x1(n) from the main noise source s1(n), and then x1(n) replaces x(n) and enters the control system to participate in the calculation.

[0036] In recent years, networks based on deep learning have performed well in the task of sound source separation, and some networks even have performance comparable to that of the human ear. These networks describe the separation task as a supervised learning problem and develop statistical models from the perspective of task optimization. Therefore, the present invention uses a network based on deep learning to construct a time domain sound source separation system.

[0037] Before ANC starts, the separation network is trained. First, a training data set is established, which should contain multiple noise sources. Two different types of sound sources are randomly selected from the training data set, and mixed with a random signal-to-noise ratio to obtain a mixed sound source. An end-to-end time-domain sound source separation system is built, and a time-domain separation network is created based on the encoder-separator-decoder framework. The encoder is implemented through a one-dimensional convolutional layer. The separator is used to estimate the masking matrix of the sound source. The decoder is a one-dimensional deconvolutional layer. The masking matrix of each sound source from the separator is multiplied by the two-dimensional expression of the mixed sound from the encoder, and then the time domain signal of the separated sound source can be synthesized through the decoder. The time domain separation network is trained using the obtained mixed sound source data to obtain an end-to-end time-domain sound source separation system. The network framework is as follows: Figure 2 shown.

[0038] The separation network has a variety of evaluation indicators, which mainly measure the performance of the network based on the difference between the signal estimated by the network and the original signal. The input of the network is the time domain signal x of the mixed sound, which is composed of the real signals t1 and t2.

[0039] x=t1+t2 (1) The output of the network is x1 and x2, which represent the estimation of the two sound sources respectively. The estimated signal x can be decomposed into:

[0040] x=t+e interf +e noise +e artif (2)

[0041] Where t represents the original signal (linear gain is allowed). The estimated signal x also contains three types of distortion: interference from other sound sources (e interf ), channel noise caused by sensor sampling (e noise ) and other reasons (such as algorithm) introduced distortion (e artif ).

[0042] The most commonly used indicator for separation systems is the scale-invariant source-to-distortion ratio (SI-SDR).

[0043]

[0044] After the training is completed, the following steps are used to perform active noise control based on sound source separation. There is no end standard set in the algorithm, but it is controlled manually:

[0045] First, a reference sensor is used to pick up a mixed reference signal x(n) emitted by the noise source, where x(n) comes from the noise source s1(n) and the noise source s2(n);

[0046] Secondly, the mixed reference signal x(n) is separated by the established and trained time domain sound source separation system to obtain the signal x1(n) from the main noise source s1(n), which is used as the input of the control filter to obtain the secondary signal y(n).

[0047]

[0048] Where w(n) is the weight coefficient of the control filter;

[0049] Afterwards, the secondary signal drives the secondary sound source to generate a cancellation signal y'(n), and the cancellation signal y'(n) and the sound pressures d1(n) and d2(n) of the primary sound field of the noise source at the error sensor are superimposed to form an error signal e(n);

[0050] In the active noise control process, the control filter weights are iterated using the signal x1(n) of the main noise source s1(n), the secondary path transfer function h(n), and the error signal e(n):

[0051] w(n+1)=w(n)-μr(n)e(n) (5)

[0052] Where μ is the step size parameter, r(n) = x1(n)*h(n), thus achieving better noise reduction effect.

[0053] The present invention is described below in conjunction with specific embodiments. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0054] (1) First, a convolutional time-domain audio separation network (ConvTasNet) is trained. The training datasets are speech from the LibriSpeech dataset, unvoiced music from the MUSAN dataset, and ambient sound from the BBC sound effects dataset. Based on the above single-source datasets, we create mixed sound segments by randomly selecting two sound source segments and mixing them with a random energy ratio between -5dB and +5dB. To avoid confusion, segments from the same speaker, the same music, and the same ambient sound source (such as the same car) are not allowed to be mixed. Random mixing generates a large overlapping sound event dataset containing 6 different types of mixed sounds (speech+speech, music+music, ambient sound+ambient sound, speech+music, speech+ambient sound, music+ambient sound). The mixed sound dataset here is limited to 180,000 segments (150 hours), of which 70% are used for training (105 hours), 20% are used for cross-validation (30 hours), and 10% are used for testing (15 hours). The system sampling rate is set to 8000Hz.

[0055] (2) The network was trained and evaluated using the above dataset, with 100 iterations. The network was implemented based on the Asteroid toolbox of pytorch. The Adam optimizer was used for training, with an initial learning rate of 0.001. If the performance of the validation set did not improve in three consecutive epochs, the learning rate was halved. The SI-SDR of the training set and validation set changes with the number of iterations are shown in Figure 2. Figure 4 shown.

[0056] (3) The transfer function used in the simulation is the transfer function measured in an ordinary room. The frequency responses of paths P1(z) and P2(z) are as follows: Figure 5 The secondary path frequency response is shown in Figure 6 shown.

[0057] (4) Traffic noise is selected as the main noise source and speech signal as the interference noise source. The primary sound field at the error microphone is controlled by the traditional method and the method proposed in the present invention respectively. The control filter length is selected as 2048. The optimal step length parameter is selected for each method by trial and error. The step length parameter of the traditional method is 2.4621, and the step length parameter of the method proposed in the present invention is selected as 7.4057. In order to evaluate the noise reduction performance of each method, the average noise reduction (ANR) is defined as

[0058]

[0059] In the formula, e(n) = [e(n), e(n-1), ..., e(n-N+1)], d(n) = [d(n), d(n-1), ..., d(n-N+1)], and in the simulation, N = 10000. The smaller the ANR, the greater the noise reduction amount and the better the noise reduction effect.

[0060] The ANR of the two methods is as follows Figure 7 As shown in the figure, it can be seen that the method proposed by the present invention achieves better noise reduction effect than the traditional method.

[0061] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.

Claims

1. An active noise control method based on sound source separation, characterized in that: The following steps are involved: Step 1: Build and train an end-to-end time-domain sound source separation system; The time-domain sound source separation system is constructed using a deep learning-based network; The time-domain sound source separation system is constructed using an encoder-separator-decoder framework; The encoder is implemented by a one-dimensional convolutional layer, the separator is used to estimate the masking matrix of the sound source, and the decoder is a one-dimensional deconvolutional layer; the masking matrix of each sound source from the separator is multiplied by the two-dimensional representation of the mixed sound from the encoder, and then the time domain signals of each sound source after separation are synthesized by the decoder; Step 2: Use the reference sensor to pick up the mixed reference signal x(n) emitted by multiple noise sources. Step 3: Use the time domain sound source separation system established in step 1 to separate the mixed reference signal x(n) to obtain the signal x1(n) from the main noise source s1(n), which is used as the input of the control filter to obtain the secondary signal y(n). y(n)=x1 T (n)w(n) Where w(n) is the weight coefficient of the control filter; Step 4: The secondary signal drives the secondary sound source to generate a cancellation signal y'(n), and the cancellation signal y'(n) and the sound pressure of the primary sound fields of the multiple noise sources at the error sensor are superimposed to form an error signal e(n); Step 5: Iterate the control filter weights using the signal x1(n) of the primary noise source s1(n), the secondary path transfer function h(n), and the error signal e(n): w(n+1)=w(n)-μr(n)e(n) Where μ is the step size parameter, r(n)=x1(n)*h(n).

2. The active noise control method based on sound source separation according to claim 1, characterized in that: The training process of the time domain sound source separation system is: First, a training data set is established, wherein the training data set includes a plurality of noise sources; two different types of sound sources are randomly selected from the training data set, and mixed at a random signal-to-noise ratio to obtain a mixed sound source; The obtained mixed sound source data is used to train the time domain separation network to obtain an end-to-end time domain sound source separation system.

3. The active noise control method based on sound source separation according to claim 2, characterized in that: During the training process of the time-domain sound source separation system, a scale-invariant signal-to-interference ratio is used as an evaluation index of the time-domain sound source separation system.

Images