RNN-based high signal-to-noise ratio speech denoising methods, devices, equipment, and media
By using an RNN-based method, the problem of harmonic filtering in high signal-to-noise ratio speech denoising is solved, and real-time, fast and coherent speech denoising processing for high signal-to-noise ratio speech is achieved, improving the efficiency and realism of speech denoising.
Patent Information
- Application Number
- CN202211326640.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-27
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2042-10-27
AI Technical Summary
High signal-to-noise ratio speech has high quality, but harmonic filtering is difficult to process, affecting speech noise reduction efficiency, and it lacks the ability to identify and process invalid blanks.
An RNN-based approach is adopted. External noise information is acquired, and fast Fourier transform and filtering are performed. An RNN network model is constructed for feature extraction and XOR processing. Gaussian white noise is used for wavelet threshold denoising. The signal-to-noise ratio is determined and cyclic judgment and filtering are performed.
It achieves real-time, fast, and coherent speech denoising processing for high signal-to-noise ratio speech, improves the realism of the optimized sound denoising and the ability to process diverse training data, and meets the requirements for continuity and real-time performance in speech processing.
Smart Images

Figure CN115691535B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of speech denoising technology, and in particular to a high signal-to-noise ratio speech denoising method, apparatus, device, and medium based on RNN. Background Technology
[0002] In the information age, digital communication systems are widely used. However, these systems are inevitably affected by external noise. Noise takes many forms, is completely random and unpredictable. Whether in military or civilian communication systems, echo noise, power frequency interference, and random noise can severely impact communication quality. Strong noise can even cause significant harm to human physical and mental health. In high-noise environments, how to stably and accurately filter out noise from down-disturbed signals and extract the original communication signal while maintaining excellent communication clarity has become an important research direction in current digital signal processing. Signal-to-noise ratio (SNR) refers to the ratio of signal to noise in an electronic device or system. The signal refers to the electronic signal from outside the device that needs to be processed by it. Noise refers to irregular extraneous signals (or information) that are not present in the original signal after passing through the device, and this noise does not change with the original signal. A higher SNR indicates that the device generates less noise. Generally speaking, a higher SNR means less noise mixed in with the signal, resulting in higher sound quality during playback, and vice versa.
[0003] During noise reduction, by performing a Fast Fourier Transform on the noise signal and plotting its spectrum, the harmonic frequency distribution of the noise is observed and analyzed. A suitable filtering module is then selected to filter out the harmonics of the noise information, thus obtaining noise-removed speech. However, since high signal-to-noise ratio speech has high quality, harmonic filtering is difficult, affecting the efficiency of speech noise reduction. Furthermore, the lack of ability to identify and process invalid blanks further affects the noise reduction capability of the signal. Summary of the Invention
[0004] The main objective of this invention is to solve the problems that, due to the high quality of high signal-to-noise ratio speech, harmonic filtering is difficult to process, which affects the efficiency of speech noise reduction, and there is a lack of ability to identify and process invalid blanks.
[0005] The first aspect of this invention provides a high signal-to-noise ratio speech denoising method based on RNN, comprising: Acquire external noise information; Based on a preset fast Fourier transform algorithm, the external noise information is processed to obtain processed information; The processing information that does not meet the preset signal-to-noise ratio value is filtered to obtain filtered information; According to the preset blank noise removal algorithm, the blank noise of the filtered information is removed to obtain clean filtered information; An RNN network model is constructed, and based on the RNN network model, the clean filtered information is subjected to feature extraction and XOR processing to obtain the filtering threshold. Based on the filtering threshold, the clean filtered information is subjected to wavelet threshold denoising using preset Gaussian white noise to obtain denoised speech. Determine whether the signal-to-noise ratio of the denoised speech meets the specified signal-to-noise ratio value; If so, then the noise reduction operation is complete; If not, the noise-reduced speech is returned to the step of filtering the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
[0006] Furthermore, in a second implementation of the first aspect of the present invention, obtaining external noise information includes: The noise spectrum in the external noise information is separated into audio features to obtain sound features with different characteristics.
[0007] Furthermore, in a third implementation of the first aspect of the present invention, the sound features include channel design, enhanced human voice frequencies, environmental noise, and instrumental sounds.
[0008] Furthermore, in a fourth implementation of the first aspect of the present invention, the step of removing blank noise from the filtered information according to a preset blank noise cancellation algorithm to obtain clean filtered information includes: The signal strength of each frame in the filtered information is detected, and a frame in the filtered information that is lower than a preset strength threshold is identified as blank noise and the blank noise is removed.
[0009] Furthermore, in a fifth implementation of the first aspect of the present invention, the construction of the RNN network model includes: Collect training set data, and collect the optimized training set data as the output layer's result value; The output value of the neuron is calculated forward, and the dimension of the threshold matrix is calculated based on the dimension of the input training set feature vector and the dimension of the output threshold direction vector. The error value of the neuron is calculated in reverse, and the error value is passed to the previous layer of the network and forward along the time line to the initial time to obtain the error term at any time. The error term is propagated backward along the timeline, and the input from the previous layer network is weighted. Based on the preset error function, partial derivatives are taken on the weighted input of the neuron; Calculate the gradient of the weights of each neuron and the gradient of the error function with respect to the weight matrix; Based on the error term and the output value of the time loop layer, calculate the gradient of the weight matrix over time to obtain the sum of the gradients at each time step within the matrix; The weights are updated based on a pre-set stochastic gradient descent algorithm; Repeat the above steps until the RNN network model is constructed.
[0010] Furthermore, in a sixth implementation of the first aspect of the present invention, the optimized training set data includes: The training set is preprocessed, and multiple feature files are extracted; The data features are obtained by performing XOR and weighted processing on multiple feature files.
[0011] Furthermore, in a seventh implementation of the first aspect of the present invention, the preprocessing includes adding noise, lowering and raising the volume, waveform shifting, waveform stretching, treble correction, and speeding up playback.
[0012] A second aspect of the present invention provides a high signal-to-noise ratio (SNR) speech denoising device based on an RNN, the RNN-based high SNR speech denoising device comprising: The acquisition module is used to acquire external noise information; The processing module is used to process the external noise information based on a preset fast Fourier transform algorithm to obtain processed information; The filtering module is used to filter the processing information that does not meet the preset signal-to-noise ratio value to obtain filtered information; The elimination module is used to remove blank noise from the filtered information according to a preset blank noise elimination algorithm, so as to obtain clean filtered information. The training module is used to construct an RNN network model and, based on the RNN network model, to extract features from the clean filtered information to obtain the filtering threshold. The noise reduction module is used to perform wavelet threshold noise reduction on the clean filtered information using preset Gaussian white noise according to the filtering threshold, so as to obtain noise-reduced speech. The comparison module is used to determine whether the signal-to-noise ratio of the denoised speech meets the signal-to-noise ratio value; The first comparison result module is used to complete the noise reduction operation if the result is yes. The second comparison result module is used to, if not, return the denoised speech to the step of filtering the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
[0013] A third aspect of the present invention provides a high signal-to-noise ratio speech denoising device based on RNN, the RNN-based high signal-to-noise ratio speech denoising device comprising: a memory and at least one processor, the memory storing instructions, and the memory and the at least one processor being interconnected via a line; The at least one processor invokes the instructions in the memory to cause the device to execute the above-described RNN-based high signal-to-noise ratio speech denoising method.
[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described RNN-based high signal-to-noise ratio speech denoising method.
[0015] The beneficial effects of this invention are as follows: By quickly determining the sound signal-to-noise ratio, interference with blank noise in the recorded sound is achieved. Furthermore, by locating the frame intervals of the audio channels, the blank noise within the intervals is quickly matched and located, satisfying the cyclic judgment processing of the RNN network. Moreover, through the designed speech noise reduction system, after the sound noise reduction process, the realism of the sound after noise reduction optimization is improved by enhancing the human voice and instrumental sounds in the remaining channels. In this embodiment, by establishing a recurrent neural network model, different features in the recorded speech can be separated. Simultaneously, the characteristics of the recurrent neural network model allow for the rapid extraction of audio noise features in the time vector of synchronously correlated speech using temporal information. This facilitates the rapid extraction of filtering noise thresholds and enables real-time, fast, and correlated speech denoising for high signal-to-noise ratio speech. This satisfies the need for continuous and real-time speech processing, and facilitates real-time speech denoising. It also eliminates interference from blank noise at the beginning and end of the recorded sound and achieves rapid matching and localization of white noise within the frame intervals of the audio channels, satisfying the recurrent judgment processing of the recurrent neural network. Furthermore, the designed speech denoising system enhances the realism of the optimized speech by amplifying the human voice and instrumental sounds in the remaining channels after denoising. Preprocessing the training set features allows for the addition of distinguishing features to the training library. During training, the feature vectors are directly extracted, expanding the training set features and effectively improving the diversity and feature-based processing of the training data, thereby enhancing training performance and improving the extraction efficiency of input features. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the first embodiment of the high signal-to-noise ratio speech denoising method based on RNN in this invention. Figure 2 This is a schematic diagram of the second embodiment of the high signal-to-noise ratio speech denoising method based on RNN in this invention. Figure 3 This is a schematic diagram of the third embodiment of the high signal-to-noise ratio speech denoising method based on RNN in this invention. Figure 4 This is a schematic diagram of the first embodiment of the high signal-to-noise ratio speech denoising device based on RNN in this invention. Figure 5 This is a schematic diagram of an embodiment of a high signal-to-noise ratio speech denoising device based on RNN in this invention. Detailed Implementation
[0017] This invention provides a high signal-to-noise ratio speech denoising method, apparatus, device, and medium based on RNN.
[0018] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1-3 One embodiment of the high signal-to-noise ratio (SNR) speech denoising method based on RNN in this invention includes: 101. Obtain external noise information; Furthermore, step 101 can also be performed as follows: 1011. The noise spectrum in the external noise information is separated into audio features to obtain sound features with different characteristics.
[0020] In this embodiment, sound features include channel design, enhanced human voice frequencies, ambient noise, and instrumental sounds.
[0021] 102. Based on a preset Fast Fourier Transform algorithm, the external noise information is processed to obtain processed information; 103. Filter the processing information that does not meet the preset signal-to-noise ratio value to obtain filtered information; 104. According to the preset blank noise removal algorithm, the blank noise of the filtered information is removed to obtain clean filtered information; 105. Construct an RNN network model, and based on the RNN network model, extract features from the clean filtered information and perform XOR processing to obtain the filtering threshold. Furthermore, step 105 can also be performed as follows: 1051. Collect training set data, and collect the optimized training set data as the result value of the output layer; 1052. Calculate the output value of the neuron forward, and calculate the dimension of the threshold matrix based on the dimension of the input training set feature vector and the dimension of the output threshold direction vector. 1053. Calculate the error value of the neuron in reverse, and pass the error value to the previous layer of the network and forward along the timeline to the initial time to obtain the error term at any time. 1054. Propagate the error term backward along the timeline and input the weighted values from the previous layer network; 1055. Based on the preset error function, perform partial derivatives on the weighted input of the neuron; 1056. Calculate the gradient of the weights of each neuron and the gradient of the error function with respect to the weight matrix; 1057. Based on the error term and the output value of the time loop layer, calculate the gradient of the weight matrix over time to obtain the sum of the gradients at each time step within the matrix; 1058. Update the weights based on a pre-set stochastic gradient descent algorithm; 1059. Repeat the above steps until the RNN network model is constructed.
[0022] Furthermore, the "optimized training set data" can also perform the following specific actions: The training set is preprocessed, and multiple feature files are extracted; The data features are obtained by performing XOR and weighted processing on multiple feature files.
[0023] The preprocessing includes adding noise, lowering and raising the volume, waveform shifting, waveform stretching, treble correction, and speeding up playback.
[0024] In this embodiment, based on the time series and the input samples, after recording the memory of the samples at a certain time, the corresponding feature weights are output according to the input weights and the feature weights of the samples. As time progresses, the aforementioned state acts as the state of the input time and participates in the prediction of the next state, thus obtaining the final weight sharing and the total past memory. Let the noise feature Q be the input value, and p be the output value and the filtering threshold; ; V is a weight matrix related to the bias rate b, and the function that identifies the hidden layer is n(1), a non-linear function, which is expressed as follows: ; Where M is the deviation vector The relevant weight matrix and activation function ( ) is using tanh.
[0025] To make p closer to the target p ,make ; The minimum error is taken after training the network.
[0026] The input values X have varying lengths and a high dimensionality. Multiple input boxes are used, and after continuous looping, the input boxes cycle continuously over time, resulting in the calculated output value p. This becomes a new input value, and the process repeats continuously, improving computational efficiency. When updating with an error value, the error loss is calculated using the cross-entropy loss function and updated using the gradient descent method, and the update method for the n(1) weight is the sum of the partial derivatives of the deviation at each update time.
[0027] 106. Based on the filtering threshold, the clean filtered information is subjected to wavelet threshold denoising using preset Gaussian white noise to obtain denoised speech; 107. Determine whether the signal-to-noise ratio of the denoised speech meets the specified signal-to-noise ratio value; 108. If yes, then the noise reduction operation is complete; 109. If not, return the noise-reduced speech to the step of filtering the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
[0028] In this embodiment, by establishing a recurrent neural network model, different features in the recorded speech can be separated. Simultaneously, the characteristics of the recurrent neural network model allow for the rapid extraction of audio noise features in the time vector of synchronously correlated speech using temporal information. This facilitates the rapid extraction of filtering noise thresholds and enables real-time, fast, and correlated speech denoising for high signal-to-noise ratio speech. This satisfies the need for continuous and real-time speech processing, and facilitates real-time speech denoising. It also eliminates interference from blank noise at the beginning and end of the recorded sound and achieves rapid matching and localization of white noise within the frame intervals of the audio channels, satisfying the recurrent judgment processing of the recurrent neural network. Furthermore, the designed speech denoising system enhances the realism of the optimized speech by amplifying the human voice and instrumental sounds in the remaining channels after denoising. Preprocessing the training set features allows for the addition of distinguishing features to the training library. During training, the feature vectors are directly extracted, expanding the training set features and effectively improving the diversity and feature-based processing of the training data, thereby enhancing training performance and improving the extraction efficiency of input features.
[0029] The high signal-to-noise ratio (SNR) speech denoising method based on RNN in the embodiments of the present invention has been described above. The high SNR speech denoising device based on RNN in the embodiments of the present invention will be described below. Please refer to [link to relevant documentation]. Figure 4 One embodiment of the high signal-to-noise ratio speech denoising device based on RNN in this invention includes: Acquisition module 201 is used to acquire external noise information; Processing module 202 is used to process the external noise information based on a preset fast Fourier transform algorithm to obtain processed information; Filtering module 203 is used to filter the processing information that does not meet the preset signal-to-noise ratio value to obtain filtered information; The elimination module 204 is used to remove blank noise from the filtered information according to a preset blank noise elimination algorithm to obtain clean filtered information. Training module 205 is used to construct an RNN network model and, based on the RNN network model, extract features from the clean filtered information to obtain a filtering threshold. The noise reduction module 206 is used to perform wavelet threshold noise reduction on the clean filtered information using preset Gaussian white noise according to the filtering threshold, so as to obtain noise-reduced speech. Comparison module 207 is used to determine whether the signal-to-noise ratio of the denoised speech meets the signal-to-noise ratio value; The first comparison result module 208 is used to complete the noise reduction operation if the result is yes. The second comparison result module 209 is used to, if not, return the noise-reduced speech to perform the filtering process on the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
[0030] In another embodiment of the high signal-to-noise ratio (SNR) speech denoising device based on RNN, the RNN-based high SNR speech denoising device includes: Acquisition module 201 is used to acquire external noise information; Processing module 202 is used to process the external noise information based on a preset fast Fourier transform algorithm to obtain processed information; Filtering module 203 is used to filter the processing information that does not meet the preset signal-to-noise ratio value to obtain filtered information; The elimination module 204 is used to remove blank noise from the filtered information according to a preset blank noise elimination algorithm to obtain clean filtered information. Training module 205 is used to construct an RNN network model and, based on the RNN network model, extract features from the clean filtered information to obtain a filtering threshold. The noise reduction module 206 is used to perform wavelet threshold noise reduction on the clean filtered information using preset Gaussian white noise according to the filtering threshold, so as to obtain noise-reduced speech. Comparison module 207 is used to determine whether the signal-to-noise ratio of the denoised speech meets the signal-to-noise ratio value; The first comparison result module 208 is used to complete the noise reduction operation if the result is yes. The second comparison result module 209 is used to, if not, return the noise-reduced speech to perform the filtering process on the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
[0031] Specifically, the acquisition module 201 shown can also execute: The noise spectrum in the external noise information is separated into audio features to obtain sound features with different characteristics, including channel design, enhanced human voice frequencies, environmental noise and instrumental sounds.
[0032] Specifically, the noise reduction module 206 can also perform the following: The signal strength of each frame in the filtered information is detected, and a frame in the filtered information that is lower than a preset strength threshold is identified as blank noise. The blank noise is then removed to obtain clean filtered information.
[0033] Specifically, the training module 205 can also perform the following: Collect training set data, and collect the optimized training set data as the output layer's result value; The output value of the neuron is calculated forward, and the dimension of the threshold matrix is calculated based on the dimension of the input training set feature vector and the dimension of the output threshold direction vector. The error value of the neuron is calculated in reverse, and the error value is passed to the previous layer of the network and forward along the time line to the initial time to obtain the error term at any time. The error term is propagated backward along the timeline, and the input from the previous layer network is weighted. Based on the preset error function, partial derivatives are taken on the weighted input of the neuron; Calculate the gradient of the weights of each neuron and the gradient of the error function with respect to the weight matrix; Based on the error term and the output value of the time loop layer, calculate the gradient of the weight matrix over time to obtain the sum of the gradients at each time step within the matrix; The weights are updated based on a pre-set stochastic gradient descent algorithm; Repeat the above steps until the RNN network model is constructed.
[0034] Specifically, the "optimized training set data" includes: The training set is preprocessed, and multiple feature files are extracted; The data features are obtained by performing XOR and weighted processing on multiple feature files.
[0035] The preprocessing includes adding noise, lowering and raising the volume, waveform shifting, waveform stretching, treble correction, and speeding up playback.
[0036] In this embodiment, by establishing a recurrent neural network model, different features in the recorded speech can be separated. Simultaneously, the characteristics of the recurrent neural network model allow for the rapid extraction of audio noise features in the time vector of synchronously correlated speech using temporal information. This facilitates the rapid extraction of filtering noise thresholds and enables real-time, fast, and correlated speech denoising for high signal-to-noise ratio speech. This satisfies the need for continuous and real-time speech processing, and facilitates real-time speech denoising. It also eliminates interference from blank noise at the beginning and end of the recorded sound and achieves rapid matching and localization of white noise within the frame intervals of the audio channels, satisfying the recurrent judgment processing of the recurrent neural network. Furthermore, the designed speech denoising system enhances the realism of the optimized speech by amplifying the human voice and instrumental sounds in the remaining channels after denoising. Preprocessing the training set features allows for the addition of distinguishing features to the training library. During training, the feature vectors are directly extracted, expanding the training set features and effectively improving the diversity and feature-based processing of the training data, thereby enhancing training performance and improving the extraction efficiency of input features.
[0037] The above is attached Figure 4 The high signal-to-noise ratio speech denoising device based on RNN in this embodiment of the invention is described in detail from the perspective of modular functional entities. The high signal-to-noise ratio speech denoising device based on RNN in this embodiment of the invention is described in detail below from the perspective of hardware processing.
[0038] Appendix Figure 5 This is a schematic diagram of a high signal-to-noise ratio (SNR) speech denoising device based on an RNN, according to an embodiment of the present invention. The RNN-based high SNR speech denoising device 300 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the RNN-based high SNR speech denoising device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the RNN-based high SNR speech denoising device 300.
[0039] The RNN-based high signal-to-noise ratio speech denoising device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 5 The illustrated structure of the RNN-based high signal-to-noise ratio speech denoising device does not constitute a limitation on the RNN-based high signal-to-noise ratio speech denoising device. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.
[0040] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the RNN-based high signal-to-noise ratio speech denoising method, apparatus, device and medium.
[0041] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0042] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0043] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A high signal-to-noise ratio speech denoising method based on RNN, characterized in that, include: Acquire external noise information; Based on a preset fast Fourier transform algorithm, the external noise information is processed to obtain processed information; The processing information that does not meet the preset signal-to-noise ratio value is filtered to obtain filtered information; According to the preset blank noise removal algorithm, the blank noise of the filtered information is removed to obtain clean filtered information; Constructing an RNN network model; wherein, constructing the RNN network model includes: collecting training set data, and collecting optimized training set data as the output layer result value; the optimized training set data includes: preprocessing the training set and extracting multiple feature files; performing XOR processing and weighted processing on the multiple feature files to obtain data features; training the RNN network model based on the data features until the RNN network model is constructed; Based on the constructed RNN network model, the clean filtered information is used to extract features to obtain noise features, and the noise features are input into the RNN network model. The RNN network model outputs XOR and weighted processed data, and the filtering threshold is obtained based on the processed data. Based on the filtering threshold, and using preset Gaussian white noise, wavelet threshold denoising is performed on the clean filtered information to obtain denoised speech; Determine whether the signal-to-noise ratio of the denoised speech meets the specified signal-to-noise ratio value; If so, then complete the noise reduction operation; If not, the noise-reduced speech is returned to the step of filtering the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
2. The high signal-to-noise ratio speech denoising method based on RNN according to claim 1, characterized in that, The acquisition of external noise information includes: The noise spectrum in the external noise information is separated into audio features to obtain sound features with different characteristics.
3. The high signal-to-noise ratio speech denoising method based on RNN according to claim 2, characterized in that, The sound characteristics include channel design, enhanced human voice frequencies, ambient noise, and instrumental sounds.
4. The high signal-to-noise ratio speech denoising method based on RNN according to claim 1, characterized in that, The step of removing blank noise from the filtered information according to a preset blank noise cancellation algorithm to obtain clean filtered information includes: The signal strength of each frame in the filtered information is detected, and a frame in the filtered information that is lower than a preset strength threshold is identified as blank noise. The blank noise is then removed to obtain clean filtered information.
5. A high signal-to-noise ratio speech denoising device based on RNN, characterized in that, include: The acquisition module is used to acquire external noise information; The processing module is used to process the external noise information based on a preset fast Fourier transform algorithm to obtain processed information; The filtering module is used to filter the processing information that does not meet the preset signal-to-noise ratio value to obtain filtered information; The elimination module is used to remove blank noise from the filtered information according to a preset blank noise elimination algorithm, so as to obtain clean filtered information. A training module is used to construct an RNN network model. The construction of the RNN network model includes: collecting training set data and collecting optimized training set data as the output layer's result value; the optimized training set data includes: preprocessing the training set and extracting multiple feature files; performing XOR and weighted processing on the multiple feature files to obtain data features; training the RNN network model based on the data features until the RNN network model is constructed; and based on the constructed RNN network model, extracting features from the clean filtering information to obtain noise features, inputting the noise features into the RNN network model, and having the RNN network model output XORed and weighted processed data, and obtaining a filtering threshold based on the processed data. The noise reduction module is used to perform wavelet threshold noise reduction on the clean filtered information based on the filtering threshold and using preset Gaussian white noise to obtain noise-reduced speech. The comparison module is used to determine whether the signal-to-noise ratio of the denoised speech meets the signal-to-noise ratio value; The first comparison result module is used to complete the noise reduction operation if the result is yes. The second comparison result module is used to, if not, return the denoised speech to the step of filtering the processing information that does not meet the preset signal-to-noise ratio value to obtain the filtered information.
6. A high signal-to-noise ratio speech denoising device based on RNN, characterized in that, The high signal-to-noise ratio speech denoising device based on RNN includes: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are interconnected via a line; The at least one processor invokes the instructions in the memory to cause the RNN-based high signal-to-noise ratio speech denoising device to perform the RNN-based high signal-to-noise ratio speech denoising method as described in any one of claims 1-4.
7. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the high signal-to-noise ratio speech denoising method based on RNN as described in any one of claims 1-4.
Citation Information
Patent Citations
Human voice recognition algorithm
CN110033757A
Long-sequence biological hash authentication method based on feature fusion
CN112967724A
