A noise suppression method and electronic equipment

Abstract

The disclosed embodiment is about a method and electronic device for noise suppression, including: downsampling a speech signal; preprocessing the speech signal within the sampling rate to divide the speech signal into multiple sub-bands; normalizing the speech data corresponding to each sub-band; inputting the speech data of each sub-band into the pre-trained target neural network model corresponding to each sub-band to obtain multiple groups of result data corresponding to each sub-band; synthesizing the multiple groups of result data to obtain the final noise suppression result. In this embodiment, by decomposing the speech signal into different sub-bands, constructing different deep learning networks for processing different sub-bands, and finally synthesizing the noise suppression results of each sub-band, it is more refined and has a better noise suppression effect.

Description

Background Art

[0002] Traditional noise suppression methods are based on frequency domain analysis or statistical features. For example, the noise suppression algorithm of the famous open source project WebRtc performs frequency domain analysis on voice signals, calculates some statistical features in the frequency domain, generates a probability model of noise, and then performs denoising through Wiener filtering. Traditional noise suppression methods work better for stationary noise, but have poor results for non-stationary noise. However, in most scenarios, our call environments are complex and accompanied by non-stationary noise. Traditional methods cannot achieve the desired results in many cases.

[0003] At present, deep learning big data technology is widely used in many fields, including noise suppression of speech signals, but deep learning methods generally build a deep learning network. But in fact, for speech signals, the information contained in different frequency bands is different, and their contribution to noise suppression is also different.

[0004] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0005] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[0006] The purpose of the embodiments of the present disclosure is to provide a noise suppression method and electronic device, thereby overcoming one or more problems caused by limitations and defects of related technologies at least to a certain extent.

[0007] The present disclosure provides a method for noise suppression, including:

[0008] Down-sampling the speech signal;

[0009] Preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands;

[0010] Performing data normalization processing on the speech data corresponding to each of the sub-frequency bands;

[0011] Inputting the speech data of each sub-frequency band into the pre-trained target neural network model corresponding to each sub-frequency band, respectively, to obtain multiple groups of result data corresponding to each sub-frequency band;

[0012] The multiple groups of result data are synthesized to obtain the final noise suppression result.

[0013] In one embodiment of the present disclosure, the step of preprocessing the speech signal within the sampling rate includes:

[0014] Pass the speech signal through a Hamming window and then perform a fast Fourier transform (FFT);

[0015] Determine the length of the FFT, and equally divide the FFT result according to the length;

[0016] The divided groups of FFT results are inversely transformed to form a plurality of sub-frequency bands.

[0017] In one embodiment of the present disclosure, the neural network model is an LSTM network model;

[0018] The target neural network model is trained in the following way:

[0019] Down-sampling the speech signal;

[0020] Preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands;

[0021] Performing data normalization processing on the speech data corresponding to each of the sub-frequency bands;

[0022] The speech data of each of the sub-bands is respectively input into the original LSTM network model corresponding to each sub-band, and each sub-band is trained to obtain the network parameters of the target LSTM network model corresponding to each sub-band, and the target LSTM network model corresponding to each sub-band is generated according to the network parameters of the target LSTM network model corresponding to each sub-band.

[0023] In one embodiment of the present disclosure, the forget gate in the LSTM network model passes through a sigmoid activation function and then passes through a preset activation function again so that the output value is between 0.5 and 1.

[0024] In one embodiment of the present disclosure, the step of performing data normalization processing on the speech data corresponding to each of the sub-frequency bands further includes:

[0025] The pulse code modulation PCM data of each sub-frequency band is normalized to a range of -1 to 1.

[0026] In one embodiment of the present disclosure, the step of obtaining multiple groups of result data corresponding to each of the sub-frequency bands includes:

[0027] Passing a plurality of groups of result data through a Hamming window and performing FFT transformation processing respectively to generate FFT results;

[0028] It is determined whether the generated FFT result is within the frequency range of the group of result data. If there are values ​​that are not within the frequency range of the group of result data, the values ​​that are not within the frequency range of the group of result data are removed.

[0029] In one embodiment of the present disclosure, the step of synthesizing the multiple groups of result data to obtain the final noise suppression result includes:

[0030] synthesizing the FFT results generated by processing the plurality of sub-frequency bands, and generating synthesized FFT data;

[0031] Performing inverse fast Fourier transform (IFFT) on the synthesized FFT data to obtain a final filtering result;

[0032] The filtering result is subjected to denormalization processing to obtain a final noise suppression result.

[0033] In one embodiment of the present disclosure, the speech signal is divided into n sub-bands, where n=2 k , where k is an integer in the range of 1 to 7.

[0034] An embodiment of the present disclosure provides an electronic device, the electronic device comprising:

[0035] A sampling module, used for down-sampling the speech signal;

[0036] A division module, used for preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands;

[0037] A normalization processing module, used for performing data normalization processing on the speech data corresponding to each of the sub-frequency bands;

[0038] A neural network processing module, used for inputting the speech data of each sub-frequency band into a pre-trained target neural network model corresponding to each sub-frequency band, so as to obtain multiple groups of result data corresponding to each sub-frequency band;

[0039] The synthesis module is used to synthesize multiple groups of result data to obtain a final noise suppression result.

[0040] An embodiment of the present disclosure provides an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps in the noise suppression method described in the above embodiment are implemented.

[0041] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0042] The noise suppression method and electronic device in the disclosed embodiment decomposes the speech signal into different sub-bands, constructs different deep learning networks for processing the different sub-bands, and finally synthesizes the noise suppression results of each sub-band, which is more detailed and has a better noise suppression effect. In addition, by constructing an improved neural network for noise suppression of speech signals in different sub-bands, the frequency domain characteristics of the speech signal are fully utilized, the noise of each frequency band can be suppressed more accurately, and problems that traditional methods cannot handle can be better handled. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0044] Figure 1 A schematic flow chart showing a noise suppression method in an exemplary embodiment of the present disclosure;

[0045] Figure 2 A block diagram of a new LSTM network in an exemplary embodiment of the present disclosure is shown;

[0046] Figure 3 A schematic diagram of an input sliding window in an exemplary embodiment of the present disclosure is shown;

[0047] Figure 4 A schematic diagram of the structure of an electronic device in an exemplary embodiment of the present disclosure is shown;

[0048] Figure 5 A schematic diagram of the structure of an electronic device in an exemplary embodiment of the present disclosure is shown.

[0049] Reference numerals:

[0050] 100, sampling module; 200, division module; 300, normalization processing module; 400, neural network processing module; 500, synthesis module; 600, electronic device; 620, storage unit; 610, processing unit; 630, bus; 640, display unit; 6201, random access memory unit (RAM) 6202, cache memory unit; 6203, read-only memory unit (ROM); 6204, program / utility; 6205, program module; 650, input / output (I / O) interface; 660, network adapter; 700, external device. DETAILED DESCRIPTION

[0051] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0052] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0053] This example embodiment provides a method for noise suppression, such as Figure 1 As shown, it may include:

[0054] Step S101, downsampling the speech signal;

[0055] Step S102, preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands;

[0056] Step S103, performing data normalization processing on the speech data corresponding to each of the sub-frequency bands;

[0057] Step S104, inputting the speech data of each sub-frequency band into the pre-trained target neural network model corresponding to each sub-frequency band, so as to obtain multiple groups of result data corresponding to each sub-frequency band;

[0058] Step S105, synthesizing the multiple groups of result data to obtain a final noise suppression result.

[0059] Specifically, before noise reduction is performed in this embodiment, a deep learning network needs to be trained. Before that, voice data preparation work needs to be performed. For example, different types of volunteers can be selected, including different age groups, different regions, different languages, and different genders. A voice is recorded in a recording studio. The recording in the recording studio is usually considered to be noise-free, that is, a pure voice signal. Then the recorded voice signal is manually divided into 60 seconds per segment as the shortest voice sequence length. Then, noise signals in various scenarios are obtained, such as fan noise, noise generated by machines, and crowd noise. These noise signals are also divided into 60 seconds per segment. Then the pure voice signal and the noise signal are superimposed to generate a noisy signal. If there are n segments of pure voice signals and m segments of noise signals, then n×m segments of noisy signals will be generated after superposition, and the corresponding pure voice signal is used as the ideal output.

[0060] To facilitate the processing of the deep learning network later, after the voice data is prepared, the data needs to be preprocessed. Generally, the following three steps can be performed. Step 1: downsample the data. For voice signals, the sampling rate can be set to 8kHz, but there is no specific restriction. In this embodiment, 8kHz is taken as an example. Step 2: sub-band division. Because the sampling rate of the voice signal is 8kHz, the highest frequency domain of the voice signal is 4kHz. The voice signal is first passed through a Hamming window, and then a fast Fourier transform FFT is performed, that is, the time domain is converted into the frequency domain, and then divided into 8 groups of FFTs according to the FFT length, and then these 8 groups of FFTs are subjected to fast Fourier inverse transform IFFT to the time domain. It should be noted that these 8 groups of data will be trained separately in the future to obtain 8 LSTM network parameters. Step 3: data normalization, normalize the PCM data of the 8 divided sub-bands to the range of -1 to 1.

[0061] like Figure 2 As shown, based on the strong correlation in the timing of the speech signal, an LSTM network is used in this embodiment. The core of LSTM is the cell state, which is represented by a horizontal line running through the cell. An LSTM contains three gates to control the cell state, namely the forget gate, input gate, and output gate. For details, please refer to the prior art for understanding, and will not be repeated here. However, the traditional LSTM forget gate passes through the sigmod activation function, and the output value of the sigmod activation function is 0~1. The forget gate determines how much information from the previous state will be discarded when it flows to the next state. If the traditional LSTM is used, after passing through multiple LSTMs, the data stream passing through the forget gate is likely to be filtered out. The improved LSTM mainly modifies the output of the forget gate. t The output value of the post-activation function is 0.5 to 1. This setting can effectively reduce information loss. The improved LSTM formula is defined as follows:

[0062]

[0063]

[0064] i t =σ(W i x t +U i h t-1 +b i ) (3)

[0065] f t =σ(W f x t +U f h t-1 +b f ) (4)

[0066] o t =σ(W o x t +U o h t-1 +b o ) (5)

[0067] g t =tanh(W g x t +U g h t-1 +b g ) (6)

[0068] F t =σ(f t ) (7)

[0069]

[0070]

[0071] Among them, c represents the internal state, h represents the system state, f represents the forget gate, i represents the input gate, o represents the output gate, σ represents the sigmoid activation function, and tanh represents the tanh activation function. The improvement of the traditional LSTM network also makes the algorithm converge faster. As the data samples gradually increase, the noise suppression results will gradually improve.

[0072] like Figure 3 As shown in the figure, each segment of speech data is 60 seconds long, and each segment of speech data corresponds to 8 sub-band speech data. t For 20 milliseconds of data, a sliding window method is adopted, so that each small grid represents a data length of 10 milliseconds, and input is performed every 10 milliseconds. For a single channel, 20 milliseconds of speech data has 160 samples. That is, the input data X t It is a 160-dimensional column vector. The improved LSTM network is applied to train the 8 sub-bands respectively, and the network parameters of the 8 improved LSTM networks are obtained respectively.

[0073] Then use the trained LSTM network model parameters for denoising, the steps are as follows: 1. Downsample the speech data. 2. Divide the speech data into sub-bands. 3. Normalize the data. 4. Input the data of the 8 sub-bands into the corresponding 8 improved LSTM networks to obtain 8 groups of data. 5. Perform FFT on these 8 groups of data through the Hamming window. That is, for the FFT results, remove the frequency components of the non-group intervals. 6. Synthesize the FFT results of these 8 groups of data. 7. Perform IFFT on the synthesized FFT to obtain the final filtering result in the time domain, and denormalize it.

[0074] In this implementation, by decomposing the speech signal into different sub-bands, constructing different deep learning networks for processing different sub-bands, and finally synthesizing the noise suppression results of each sub-band, it is more refined and has a better noise suppression effect. In addition, by constructing an improved neural network for noise suppression of speech signals in different sub-bands, the frequency domain characteristics of the speech signal are fully utilized, the noise of each frequency band can be suppressed more accurately, and problems that traditional methods cannot handle can be better handled.

[0075] Optionally, in step S102, it also includes step S1021, allowing the speech signal to pass through a Hamming window and then perform a fast Fourier transform FFT; step S1022, determining the length of the FFT and equally dividing the FFT results according to the length; step S1023, inverse transforming the divided multiple groups of FFT results to form multiple sub-bands.

[0076] Specifically, because the speech sampling rate is 8kHz, the highest frequency domain of the speech signal is 4kHz. The speech signal first passes through the Hamming window and then performs FFT. The FFT length can be 128, which is not limited here. Therefore, the FFT energy values ​​of 128 frequency points can be obtained, and these 128 frequency point FFT energy values ​​can be divided into 8 groups, such as 0-15, 16-31, 32-47, ..., 112-127. Thus, the speech signal is divided into 8 sub-bands.

[0077] Optionally, in some embodiments, the neural network model is an LSTM network model; the target neural network model is trained in the following manner: step S601, downsampling the speech signal; step S602, preprocessing the speech signal within the sampling rate to divide the speech signal into multiple sub-bands; step S603, normalizing the speech data corresponding to each of the sub-bands; step S604, inputting the speech data of each of the sub-bands into the original LSTM network model corresponding to each sub-band, and training each sub-band to obtain the network parameters of the target LSTM network model corresponding to each sub-band, and generating the target LSTM network model corresponding to each sub-band according to the network parameters of the target LSTM network model corresponding to each sub-band.

[0078] Specifically, the first three steps of training the target neural network model are the same as the steps of performing noise suppression, which can be understood by referring to the above embodiment. Figure 3 As shown in the figure, each segment of speech data is 60 seconds long, and each segment of speech data corresponds to speech data of 8 sub-bands. tFor 20 milliseconds of data, a sliding window method is adopted, so that each small grid represents a data length of 10 milliseconds, and input is performed every 10 milliseconds. For a single channel, 20 milliseconds of speech data has 160 samples. That is, the input data X t It is a 160-dimensional column vector. The improved LSTM network is applied to train the 8 sub-bands respectively, and the network parameters of the 8 improved LSTM networks are obtained respectively.

[0079] Optionally, in some embodiments, the forget gate in the LSTM network model passes through the sigmod activation function and then passes through the preset activation function again so that the output value is between 0.5 and 1. Specifically, the traditional LSTM forget gate passes through the sigmod activation function, and the output value of the sigmod activation function is 0 to 1. The forget gate determines how much information from the previous state will be discarded when it flows to the next state. If the traditional LSTM is used, after passing through multiple LSTMs, the data flow passing through the forget gate is likely to be filtered out. The improved LSTM mainly modifies the output of the forget gate. After F t The output value of the post-activation function is 0.5 to 1. This setting can effectively reduce information loss.

[0080] Optionally, in some embodiments, step S103 includes step S1032, normalizing the pulse code modulation PCM data of each sub-band to the interval of -1 to 1. Specifically, pulse code modulation is to transform a time-continuous and continuous analog signal into a time-discrete and discrete digital signal and transmit it in the channel. Pulse code modulation is the process of sampling the analog signal first, then quantizing the sample amplitude, and encoding it. It can be simply understood as the process of analog / digital conversion.

[0081] Optionally, in some embodiments, step S104 includes step S1041, so that multiple groups of result data pass through a Hamming window and are respectively subjected to FFT transformation processing to generate FFT results; step S1042, judging whether the generated FFT results are within the frequency range of the group of result data; if there are values ​​that are not within the frequency range of the group of result data, removing the values ​​that are not within the frequency range of the group of result data.

[0082] Specifically, in combination with the above embodiment, the eight groups of data are subjected to a Hamming window and FFT transformation is performed respectively. For the FFT result, the frequency components not in the group interval are eliminated. For example, for the LSTM network output data of the first group, since the corresponding frequency range is 0 to 500 Hz, if there are values ​​not in this frequency after FFT transformation, they need to be eliminated.

[0083] Optionally, in some embodiments, step S105 includes step S1051, synthesizing the FFT results generated by processing the multiple sub-bands and generating synthesized FFT data; step S1052, performing inverse fast Fourier transform IFFT on the synthesized FFT data to obtain a final filtering result; and step S1053, performing denormalization on the filtering result to obtain a final noise suppression result.

[0084] Specifically, the FFT results of the eight groups of data are synthesized, and IFFT is performed on the synthesized FFT to obtain the final filtering result in the time domain, and then denormalized to obtain the final denoised speech signal.

[0085] Optionally, in some embodiments, the speech signal is divided into n sub-bands, where n=2 k , where k is an integer in the range of 1 to 7. Specifically, the above takes the example of dividing the voice data into 8 sub-bands to illustrate this embodiment. Of course, the voice data can also be divided into other numbers, and the number of divisions n=2 k , where k is an integer in the range of 1 to 7.

[0086] This example embodiment also provides an electronic device, such as Figure 4 As shown, the electronic device includes a sampling module 100, a division module 200, a normalization processing module 300, a neural network processing module 400 and a synthesis module 500.

[0087] The sampling module 100 is used to downsample the speech signal; the division module 200 is used to pre-process the speech signal within the sampling rate to divide the speech signal into multiple sub-bands; the normalization processing module 300 is used to perform data normalization processing on the speech data corresponding to each of the sub-bands; the neural network processing module 400 is used to input the speech data of each of the sub-bands into the pre-trained target neural network model corresponding to each of the sub-bands to obtain multiple groups of result data corresponding to each of the sub-bands; the synthesis module 500 is used to synthesize multiple groups of result data to obtain the final noise suppression result.

[0088] The electronic device provided in this embodiment applies the noise suppression method provided in the above embodiment. The specific implementation manner can be understood with reference to the above embodiment, which will not be described in detail here.

[0089] Optionally, in some embodiments, the division module 200 includes: a transformation module, which allows the speech signal to pass through a Hamming window and then perform a fast Fourier transform FFT; a division sub-module, which is used to determine the length of the FFT result and equally divide the FFT result according to the length; and an inverse transformation module, which is used to inverse transform the multiple groups of FFTs after division to form multiple sub-frequency bands.

[0090] Optionally, in some embodiments, the neural network model is an LSTM network model. The modules used for training the target neural network model include: a sampling submodule for downsampling the speech signal; a division submodule for preprocessing the speech signal within the sampling rate to divide the speech signal into multiple sub-bands; a normalization processing submodule for performing data normalization processing on the speech data corresponding to each of the sub-bands; a model training module for inputting the speech data of each sub-band into the original LSTM network model corresponding to each sub-band, and training each sub-band to obtain the network parameters of the target LSTM network model corresponding to each sub-band, and generating the target LSTM network model corresponding to each sub-band according to the network parameters of the target LSTM network model corresponding to each sub-band.

[0091] Optionally, in some embodiments, the normalization processing module 300 further includes: a normalization submodule, configured to normalize the pulse code modulation PCM data of each of the sub-frequency bands to a range of -1 to 1.

[0092] Optionally, in some embodiments, the neural network processing module 400 includes: a generation module, which allows multiple groups of result data to pass through a Hamming window and perform FFT transformation processing respectively to generate FFT results; a judgment module, which is used to judge whether the generated FFT results are within the frequency range of the group of result data, and if there are values ​​that are not within the frequency range of the group of result data, then remove the values ​​that are not within the frequency range of the group of result data.

[0093] Optionally, in some embodiments, the synthesis module 500 includes: a synthesis sub-module, used to synthesize the FFT results generated by processing the multiple sub-bands and generate synthesized FFT data; a filtering module, used to perform inverse fast Fourier transform IFFT on the synthesized FFT data to obtain a final filtering result; and a denormalization module, used to perform denormalization on the filtering result to obtain a final noise suppression result.

[0094] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0095] It should be noted that although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiment of the present disclosure, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of a module or unit described above can be further divided into multiple modules or units for concretization. The components displayed as modules or units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed scheme. Those of ordinary skill in the art can understand and implement it without paying creative work.

[0096] In an exemplary embodiment of the present disclosure, an electronic device is further provided, which may include a processor and a memory for storing executable instructions of the processor, wherein the processor is configured to execute the steps of the noise suppression method in any of the above embodiments by executing the executable instructions.

[0097] It will be appreciated by those skilled in the art that various aspects of the present invention may be implemented as a system, method or program product. Therefore, various aspects of the present invention may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to herein as a "circuit", "module" or "system".

[0098] Refer to the following Figure 5 The electronic device 600 according to this embodiment of the present invention is described. Figure 5 The electronic device 600 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0099] like Figure 5 As shown, the electronic device 600 is in the form of a general computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different system components (including the storage unit 620 and the processing unit 610), a display unit 640, etc.

[0100] The storage unit stores a program code, which can be executed by the processing unit 610, so that the processing unit 610 performs the steps according to various exemplary embodiments of the present invention described in the above noise suppression method section of this specification. For example, the processing unit 610 can perform the following steps: Figure 1 Follow the steps shown in .

[0101] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 6201 and / or a cache memory unit 6202 , and may further include a read-only memory unit (ROM) 6203 .

[0102] The storage unit 620 may also include a program / utility 6204 having a set (at least one) of program modules 6205, such program modules 6205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include the implementation of a network environment.

[0103] Bus 630 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0104] The electronic device 600 may also communicate with one or more external devices 700 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 600, and / or communicate with any device that enables the electronic device 600 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 650. Furthermore, the electronic device 600 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 660. The network adapter 660 may communicate with other modules of the electronic device 600 via a bus 630. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0105] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, or a network device, etc.) to perform the above-mentioned noise suppression method according to the implementation of the present disclosure.

[0106] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A method for noise suppression, characterized in that: include: Down-sampling the speech signal; Preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands; Performing data normalization processing on the speech data corresponding to each of the sub-frequency bands; Inputting the speech data of each sub-frequency band into the pre-trained target neural network model corresponding to each sub-frequency band, respectively, to obtain multiple groups of result data corresponding to each sub-frequency band; Synthesizing the multiple groups of result data to obtain a final noise suppression result; The target neural network model is trained in the following way: Down-sampling the speech signal; Preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands; Performing data normalization processing on the speech data corresponding to each of the sub-frequency bands; Inputting the speech data of each of the sub-bands into the original LSTM network model corresponding to each sub-band, respectively, and training each sub-band to obtain the network parameters of the target LSTM network model corresponding to each sub-band, and generating the target LSTM network model corresponding to each sub-band according to the network parameters of the target LSTM network model corresponding to each sub-band; The target LSTM network model is an improved LSTM network model, and the calculation formula of the improved LSTM network model is as follows: i t =σ(W i X t +U i h t-1 +b i )(3) f t =σ(W f X t +U f h t-1 +b f )(4) the t =σ(W o X t +U o h t-1 +b o (5) g t =tanh(W g X t +U g h t-1 +b g )(6) F t =σ(f t (7) Among them, c represents the internal state, h represents the system state, f represents the forget gate, i represents the input gate, o represents the output gate, σ represents the sigmoid activation function, and tanh represents the tanh activation function; The forget gate in the LSTM network model passes through the sigmoid activation function and then passes through the preset activation function again so that the output value is between 0.5 and 1.

2. The method for noise suppression according to claim 1, characterized in that: The step of preprocessing the speech signal within the sampling rate comprises: Pass the speech signal through a Hamming window and then perform a fast Fourier transform (FFT); Determine the length of the FFT, and equally divide the FFT result according to the length; The divided groups of FFT results are inversely transformed to form a plurality of sub-frequency bands.

3. The method for noise suppression according to claim 1, characterized in that: The step of performing data normalization processing on the speech data corresponding to each of the sub-frequency bands also includes: The pulse code modulation PCM data of each sub-frequency band is normalized to a range of -1 to 1.

4. The method for noise suppression according to claim 1, characterized in that: The step of obtaining multiple groups of result data corresponding to each of the sub-frequency bands comprises: Passing a plurality of groups of result data through a Hamming window and performing FFT transformation processing respectively to generate FFT results; It is determined whether the generated FFT result is within the frequency range of the group of result data. If there are values ​​that are not within the frequency range of the group of result data, the values ​​that are not within the frequency range of the group of result data are removed.

5. The method for noise suppression according to claim 4, characterized in that: The step of synthesizing the multiple groups of result data to obtain the final noise suppression result includes: synthesizing the FFT results generated by processing the plurality of sub-frequency bands, and generating synthesized FFT data; Performing inverse fast Fourier transform (IFFT) on the synthesized FFT data to obtain a final filtering result; The filtering result is subjected to denormalization processing to obtain a final noise suppression result.

6. The method for noise suppression according to claim 1, characterized in that: The speech signal is divided into n sub-bands, where n=2 k , where k is an integer in the range of 1 to 7.

7. An electronic device, characterized in that: The electronic device applies the noise suppression method according to any one of claims 1 to 6, and the electronic device comprises: A sampling module, used for down-sampling the speech signal; A division module, used for preprocessing the speech signal within the sampling rate to divide the speech signal into a plurality of sub-bands; A normalization processing module, used for performing data normalization processing on the speech data corresponding to each of the sub-frequency bands; A neural network processing module, used for inputting the speech data of each sub-frequency band into a pre-trained target neural network model corresponding to each sub-frequency band, so as to obtain multiple groups of result data corresponding to each sub-frequency band; The synthesis module is used to synthesize multiple groups of result data to obtain a final noise suppression result.

8. An electronic device, characterized in that: include: A memory, a processor, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the steps in the noise suppression method according to any one of claims 1 to 6 are implemented.

Images