CN112669870A - Training method and device of speech enhancement model and electronic equipment - Google Patents
Training method and device of speech enhancement model and electronic equipment Download PDFInfo
- Publication number
- CN112669870A CN112669870A CN202011549088.4A CN202011549088A CN112669870A CN 112669870 A CN112669870 A CN 112669870A CN 202011549088 A CN202011549088 A CN 202011549088A CN 112669870 A CN112669870 A CN 112669870A
- Authority
- CN
- China
- Prior art keywords
- weight
- voice
- calculating
- speech
- enhancement model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 91
- 238000012549 training Methods 0.000 title claims abstract description 76
- 230000000873 masking effect Effects 0.000 claims abstract description 52
- 230000000694 effects Effects 0.000 claims abstract description 42
- 230000008569 process Effects 0.000 claims abstract description 22
- 238000002372 labelling Methods 0.000 claims description 20
- 238000004364 calculation method Methods 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 9
- 238000001228 spectrum Methods 0.000 description 17
- 238000010586 diagram Methods 0.000 description 16
- 230000006870 function Effects 0.000 description 15
- 239000013598 vector Substances 0.000 description 8
- 238000004590 computer program Methods 0.000 description 7
- 238000004891 communication Methods 0.000 description 6
- 230000003287 optical effect Effects 0.000 description 6
- 238000000605 extraction Methods 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 239000000126 substance Substances 0.000 description 3
- 230000001629 suppression Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 230000014759 maintenance of location Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 239000013307 optical fiber Substances 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 239000004065 semiconductor Substances 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Landscapes
- Filters That Use Time-Delay Elements (AREA)
- Electrically Operated Instructional Devices (AREA)
Abstract
The embodiment of the disclosure discloses a training method and device of a speech enhancement model, electronic equipment and a computer-readable storage medium. The training method of the speech enhancement model comprises the following steps: acquiring a training data set; acquiring a weight parameter; inputting the characteristics of the voice with noise into a voice enhancement model to obtain prediction time-frequency masking; calculating a first loss value; calculating a weight value according to the weight parameter; calculating a second loss value according to the weight value and the first loss value; updating the model parameters of the voice enhancement model according to the second loss value; and iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached. According to the method, the updating process of the model parameters is influenced through the weight coefficients, so that the voice enhancement model achieves the preset effect, and the technical problem that the voice enhancement effect cannot be controlled when the voice enhancement model is trained in the prior art is solved.
Description
Technical Field
The present disclosure relates to the field of speech processing, and in particular, to a method and an apparatus for training a speech enhancement model, an electronic device, and a computer-readable storage medium.
Background
Currently, the speech enhancement techniques mainly include two types: a signal-based speech enhancement method and a deep learning-based speech enhancement method. In the case of computer-aided speech enhancement, signal-based methods are mostly used, which are generally relatively computationally intensive and are less robust against non-stationary noise. Later, as computer performance improved, speech enhancement methods based on deep learning developed rapidly. The method can achieve a good suppression effect on both steady-state noise and non-steady-state noise.
In the speech enhancement method based on deep learning, the learning objectives are divided into the following parts: time-frequency mask (mask) based and time-frequency spectrum based methods. Based on the mask method, when the deep learning model is trained, the mask value is used as a target, and target information is easier to learn; the method based on the time-frequency spectrum can directly use the wanted speech spectrum as the target when training the deep learning model, and can more directly learn the time-frequency information of the target speech.
However, the mask-based approach may make it easier for the model to learn the target information, but ignores the absolute energy information in the time-frequency spectrum; although the method based on the time spectrum can learn the energy information of the time spectrum more directly, the learning process is difficult, and a better convergence effect is not easy to achieve. And, they all have a problem in that the effect of speech enhancement cannot be controlled.
Disclosure of Invention
This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.
In order to solve the above technical problem, in a first aspect, an embodiment of the present disclosure provides a method for training a speech enhancement model, including:
acquiring a training data set, wherein the training data set comprises noisy voices and corresponding marked time-frequency masks;
obtaining a weight parameter, wherein the weight parameter is used for adjusting the effect of the voice enhancement model;
inputting the characteristics of the voice with noise into a voice enhancement model to obtain prediction time-frequency masking;
calculating a first loss value according to the labeling time-frequency masking and the prediction time-frequency masking;
calculating a weight value according to the weight parameter;
calculating a second loss value according to the weight value and the first loss value;
updating the model parameters of the voice enhancement model according to the second loss value;
and iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached.
Further, the calculating a weight value according to the weight parameter includes:
calculating a weight value according to the weight parameter and the labeling time-frequency masking; alternatively, the first and second electrodes may be,
calculating a weight value according to the weight parameter and the characteristics of the voice with noise; alternatively, the first and second electrodes may be,
and calculating the weight value according to the weight parameter, the characteristics of the voice with noise and the marking time-frequency mask.
Further, the calculating a weight value according to the weight parameter and the labeling time-frequency mask includes:
according to
Calculating a weight value;
wherein weight represents the weight value, and label represents the labeling time-frequency masking; a, c, d and beta represent the weight parameters, A represents the gain of the weight values, and c, d and beta are used for translating and scaling label, wherein A, d and beta are positive numbers, and c is an arbitrary number.
Further, the calculating a weight value according to the weight parameter and the characteristic of the noisy speech includes:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; a, a, b and alpha represent the weight parameters, A represents the gain of the weight values, and a, b and alpha are used for translating and scaling the data _ in, wherein A, b and alpha are positive numbers, and a is an arbitrary number.
Further, the calculating a weight value according to the weight parameter and the characteristic of the noisy speech includes:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; label represents the annotation time-frequency masking; a, a, b, c, d, alpha and beta represent the weight parameters, A represents the gain of the weight values, a, b and alpha are used for translating and scaling data _ in, c, d and beta are used for translating and scaling label, wherein A, b, d, alpha and beta are positive numbers, and a and c are arbitrary numbers.
Further, the calculating a second loss value according to the weight value and the first loss value includes:
calculating a Hadamard product of the weight value and the first loss value as a second loss value.
In a second aspect, an embodiment of the present disclosure provides a method for controlling a speech enhancement effect, including:
receiving an input signal of a weight parameter to obtain the weight parameter, wherein the weight parameter corresponds to a target voice enhancement effect;
obtaining a speech enhancement model according to the weight parameters by using the training method of the speech enhancement model according to the first aspect;
inputting the voice to be enhanced into the voice enhancement model to obtain prediction time-frequency masking;
and obtaining the enhanced voice according to the voice to be enhanced and the predicted time-frequency masking.
In a third aspect, an embodiment of the present disclosure provides a training apparatus for a speech enhancement model, including:
the training set acquisition module is used for acquiring a training data set, wherein the training data set comprises noisy voices and corresponding marked time-frequency masks;
the weight parameter acquisition module is used for acquiring weight parameters, and the weight parameters are used for adjusting the effect of the voice enhancement model;
the prediction module is used for inputting the characteristics of the noisy speech into a speech enhancement model to obtain prediction time-frequency masking;
a first loss calculation module, configured to calculate a first loss value according to the labeling time-frequency mask and the prediction time-frequency mask;
the weight calculation module is used for calculating a weight value according to the weight parameter;
a second loss calculation module, configured to calculate a second loss value according to the weight value and the first loss value;
the parameter updating module is used for updating the model parameters of the voice enhancement model according to the second loss value;
and the iteration module is used for iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached.
Further, the weight calculation module is further configured to:
calculating a weight value according to the weight parameter and the labeling time-frequency masking; alternatively, the first and second electrodes may be,
calculating a weight value according to the weight parameter and the characteristics of the voice with noise; alternatively, the first and second electrodes may be,
and calculating the weight value according to the weight parameter, the characteristics of the voice with noise and the marking time-frequency mask.
Further, the weight calculation module is further configured to:
according to
Calculating a weight value;
wherein weight represents the weight value, and label represents the labeling time-frequency masking; a, c, d and beta represent the weight parameters, A represents the gain of the weight values, and c, d and beta are used to translate and scale label, where A, c, d and beta are positive numbers.
Further, the weight calculation module is further configured to:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; a, a, b and alpha represent the weight parameters, A represents the gain of the weight values, and a, b and alpha are used for translating and scaling the data _ in, wherein A, a, b and alpha are positive numbers.
Further, the weight calculation module is further configured to:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; label represents the annotation time-frequency masking; a, a, b, c, d, alpha and beta represent the weight parameters, A represents the gain of the weight values, a, b and alpha are used for translating and scaling data _ in, and c, d and beta are used for translating and scaling label, wherein A, a, b, c, d, alpha and beta are positive numbers.
Further, the second loss calculating module is further configured to:
calculating a Hadamard product of the weight value and the first loss value as a second loss value.
In a fourth aspect, an embodiment of the present disclosure provides a device for controlling a speech enhancement effect, including:
the system comprises a weight parameter receiving module, a weight parameter processing module and a voice processing module, wherein the weight parameter receiving module is used for receiving an input signal of a weight parameter to obtain the weight parameter, and the weight parameter corresponds to a target voice enhancement effect;
a training module, configured to obtain a speech enhancement model according to the weight parameter by using the training method of the speech enhancement model in the first aspect;
the input module is used for inputting the voice to be enhanced into the voice enhancement model to obtain prediction time-frequency masking;
and the enhancement module is used for obtaining enhanced voice according to the voice to be enhanced and the prediction time-frequency masking.
In a fifth aspect, an embodiment of the present disclosure provides an electronic device, including: at least one processor; and the number of the first and second groups,
a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform any of the methods described above.
In a sixth aspect, the disclosed embodiments provide a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions for causing a computer to perform the method of any one of the preceding.
The embodiment of the disclosure discloses a training method and device of a speech enhancement model, electronic equipment and a computer-readable storage medium. The training method of the speech enhancement model comprises the following steps: acquiring a training data set; acquiring a weight parameter; inputting the characteristics of the voice with noise into a voice enhancement model to obtain prediction time-frequency masking; calculating a first loss value; calculating a weight value according to the weight parameter; calculating a second loss value according to the weight value and the first loss value; updating the model parameters of the voice enhancement model according to the second loss value; and iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached. According to the method, the updating process of the model parameters is influenced through the weight coefficients, so that the voice enhancement model achieves the preset effect, and the technical problem that the voice enhancement effect cannot be controlled when the voice enhancement model is trained in the prior art is solved.
The foregoing is a summary of the present disclosure, and for the purposes of promoting a clear understanding of the technical means of the present disclosure, the present disclosure may be embodied in other specific forms without departing from the spirit or essential attributes thereof.
Drawings
The above and other features, advantages and aspects of various embodiments of the present disclosure will become more apparent by referring to the following detailed description when taken in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numbers refer to the same or similar elements. It should be understood that the drawings are schematic and that elements and features are not necessarily drawn to scale.
FIG. 1 is a schematic flow chart diagram illustrating a method for training a speech enhancement model according to an embodiment of the present disclosure;
FIG. 2 is a logic block diagram of a method of training a speech enhancement model according to an embodiment of the present disclosure;
fig. 3 is a flowchart illustrating a method for controlling a voice enhancement effect according to an embodiment of the present disclosure;
FIG. 4 is a schematic structural diagram of an embodiment of a training apparatus for a speech enhancement model provided in an embodiment of the present disclosure;
fig. 5 is a schematic structural diagram of an embodiment of a control apparatus for a speech enhancement effect according to an embodiment of the present disclosure;
fig. 6 is a schematic structural diagram of an electronic device provided according to an embodiment of the present disclosure.
Detailed Description
Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure are shown in the drawings, it is to be understood that the present disclosure may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but rather are provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the disclosure are for illustration purposes only and are not intended to limit the scope of the disclosure.
It should be understood that the various steps recited in the method embodiments of the present disclosure may be performed in a different order, and/or performed in parallel. Moreover, method embodiments may include additional steps and/or omit performing the illustrated steps. The scope of the present disclosure is not limited in this respect.
The term "include" and variations thereof as used herein are open-ended, i.e., "including but not limited to". The term "based on" is "based, at least in part, on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Relevant definitions for other terms will be given in the following description.
It should be noted that the terms "first", "second", and the like in the present disclosure are only used for distinguishing different devices, modules or units, and are not used for limiting the order or interdependence relationship of the functions performed by the devices, modules or units.
It is noted that references to "a", "an", and "the" modifications in this disclosure are intended to be illustrative rather than limiting, and that those skilled in the art will recognize that "one or more" may be used unless the context clearly dictates otherwise.
The names of messages or information exchanged between devices in the embodiments of the present disclosure are for illustrative purposes only, and are not intended to limit the scope of the messages or information.
Fig. 1 is a flowchart of an embodiment of a training method of a speech enhancement model provided in this embodiment of the present disclosure, where the training method of the speech enhancement model provided in this embodiment may be executed by a training apparatus of the speech enhancement model, the training apparatus of the speech enhancement model may be implemented as software, or implemented as a combination of software and hardware, and the training apparatus of the speech enhancement model may be integrated in a certain device in a training system of the speech enhancement model, such as a training server of the speech enhancement model or a training terminal device of the speech enhancement model. As shown in fig. 1, the method comprises the steps of:
step S101, a training data set is obtained, wherein the training data set comprises noisy voices and corresponding marked time-frequency masks;
the training data set comprises a noisy speech and a data pair marked with a time-frequency mask corresponding to the noisy speech, namely, a noisy speech signal corresponds to a marked time-frequency mask, wherein the marked time-frequency mask represents a ratio of a pure speech signal to the noisy speech signal in the noisy speech, and optionally, the ratio represents a mask between a value obtained by dividing a pure speech frequency spectrum in the noisy speech by the noisy speech frequency spectrum and is [0,1 ]. Optionally, when the noisy speech in the training data set is synthesized by pure speech and noise, and a labeled time-frequency mask corresponding to the noisy speech is obtained, firstly, performing video decomposition on the pure speech signal and the noisy speech signal to obtain a time-frequency spectrum of the pure speech signal and a time-frequency spectrum of the noisy speech, then, respectively sampling the time-frequency spectrum of the pure speech signal and the time-frequency spectrum of the noisy speech according to a preset sampling rate, dividing a value of a sampling point of the time-frequency spectrum of the pure speech signal by a value of a corresponding sampling point of the time-frequency spectrum of the noisy speech to obtain a vector, where a value of each bit in the vector is [0,1], and the vector is the time-frequency labeled mask of the noisy speech.
Returning to fig. 1, the training method of the speech enhancement model further includes, in step S102, obtaining a weight parameter, where the weight parameter is used to adjust the effect of the speech enhancement model;
the weight parameters are non-learning parameters, that is, the weight parameters are determined before the speech enhancement model is trained and are not updated with the training of the model. Optionally, the weight parameter is obtained through a human-computer interface, for example, the weight parameter input by the user is obtained through the human-computer interface. The weight parameter is used for subsequently calculating a weight value to adjust the enhancement effect of the voice enhancement model.
Returning to fig. 1, the method for training the speech enhancement model further includes step S103, inputting the characteristics of the noisy speech into the speech enhancement model to obtain a predicted time-frequency mask;
it is understood that, before the step S103, the method further includes: and extracting the characteristics of the voice with the noise. The noisy speech signal can be represented by a time domain diagram of the noisy speech, that is, a waveform diagram of the noisy speech in a time domain, the time-frequency decomposition is performed on the waveform diagram to obtain a time-frequency spectrum of the noisy speech, and then the time-frequency spectrum of the noisy speech is subjected to feature extraction to obtain features of the noisy speech.
From the basic unit of feature extraction, the features are divided into features at time frequency unit level and features at frame level. The characteristics of the time frequency unit level are extracted from a time frequency unit signal, the characteristic particles are fine, fine details can be captured, but the representation of the integrity and the global property of the voice signal is lacked, and the spatial structure and the time sequence correlation of the voice signal cannot be captured. The frame-level features are extracted from each frame signal, are large in particles and can well represent the space-time structure and the time sequence correlation of the voice signals. Compared with the characteristics of the time-frequency unit level, the method has more integrity and globality, and can represent the perceivable voice characteristics. Such features are mainly used in speech enhancement tasks that model in units of frames. To capture more context information, these systems typically combine features of neighboring frames into an input to predict an enhancement target for an entire frame. Speech enhancement features commonly used at the present stage are roughly classified into three categories: gammatone domain features, fourier transform domain features, and pitch-based features. The features of the Gamma domain are commonly used mainly for MFCC, RASTA-PLP, AMS, Gamma Feature (GF), and Multi-Resolution Cochleagram Feature (MRCG). Pitch-based features (Pitch-based features) are features at the time-frequency unit level that compute Pitch features for each time-frequency unit, representing the likelihood that the time-frequency unit is dominated by the target speech. The Fourier transform domain is commonly characterized by FFT-magnitude and FFT-log-magnitude. And framing the input signal, performing STFT on each frame of signal to obtain an STFT coefficient, and performing modulus on the coefficient to obtain the FFT-map. In order to highlight the high-frequency signal, taking the logarithm after taking the modulus to obtain the FFT-log-magnet.
The types of the features are only examples, and any feature type and feature extraction method can be used in the present disclosure, which is not limited in the present disclosure and is not repeated herein.
After obtaining the characteristics of the noisy speech, inputting the noisy speech into a speech enhancement model, where model parameters of the speech enhancement model are initialized in advance, for example, the initialization of the model parameters may be to randomly set the model parameters or to set the model parameters to preset initial parameters. Then, the model calculates the characteristics of the input voice with noise according to the model parameters to obtain a prediction time-frequency mask, namely, the characteristics of the input voice with noise are converted into the time-frequency mask through the voice enhancement model, and if the voice enhancement model is a trained model, the prediction time-frequency mask is the same as or similar to the labeling time-frequency mask. In the training phase, the predicted time-frequency masking and the labeled time-frequency masking are generally far from each other.
Returning to fig. 1, the training method of the speech enhancement model further includes, in step S104, calculating a first loss value according to the labeling time-frequency mask and the prediction time-frequency mask;
for the speech enhancement model, a loss function is preset, and the loss function is used for calculating the difference between the labeling time-frequency masking and the prediction time-frequency masking. The training of the model should minimize the expected loss function:
JEL＝E(J(W,b；o,y)) (1)
wherein J (W, b; o, y) is a loss function, { W, b } is a model parameter of the speech enhancement model, o is a characteristic of the noisy speech, and y is a labeled time-frequency mask of the noisy speech. Illustratively, the loss function may be a minimum mean square error function:
wherein the content of the first and second substances,
wherein M is the number of data samples in the training set, M is the number of the current data sample, vLFor increasing speechAnd d, outputting prediction time-frequency masking by the strong model, wherein y is the marked time-frequency masking of the voice with noise. Thus, the first loss value between the predicted time-frequency mask and the labeled time-frequency mask obtained by inputting the characteristics of the noisy speech each time can be calculated by the formula (3).
It should be understood that the above-mentioned manner for calculating the first loss value is only an example, and practically any other manner for calculating the first loss value may be applied to the embodiments of the present disclosure, and is not described herein again.
Returning to fig. 1, the training method of the speech enhancement model further includes, in step S105, calculating a weight value according to the weight parameter;
the weighted value obtained according to different weighted parameters may include information of a target voice portion of the input noisy voice, and thus the first loss value may be adjusted so that the training of the model is trained toward a predetermined effect.
Optionally, the step S105 includes:
calculating a weight value according to the weight parameter and the labeling time-frequency masking; alternatively, the first and second electrodes may be,
calculating a weight value according to the weight parameter and the characteristics of the voice with noise; alternatively, the first and second electrodes may be,
and calculating the weight value according to the weight parameter, the characteristics of the voice with noise and the marking time-frequency mask.
Optionally, the calculating a weight value according to the weight parameter and the labeling time-frequency mask includes:
according to
Calculating a weight value;
wherein weight represents the weight value, and label represents the labeling time-frequency masking; a, c, d and beta represent the weight parameters, A represents the gain of the weight values, and c, d and beta are used for translating and scaling label, wherein A, d and beta are positive numbers, and c is an arbitrary number.
Optionally, the calculating a weight value according to the weight parameter and the characteristic of the noisy speech includes:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; a, a, b and alpha represent the weight parameters, A represents the gain of the weight values, and a, b and alpha are used for translating and scaling the data _ in, wherein A, b and alpha are positive numbers, and a is an arbitrary number.
Optionally, the calculating a weight value according to the weight parameter, the characteristic of the noisy speech, and the labeled time-frequency mask includes:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; label represents the annotation time-frequency masking; a, a, b, c, d, alpha and beta represent the weight parameters, A represents the gain of the weight values, a, b and alpha are used for translating and scaling data _ in, c, d and beta are used for translating and scaling label, wherein A, b, d, alpha and beta are positive numbers, and a and c are arbitrary numbers.
The weight parameters a, b, c, d, alpha and beta are used for translating and scaling the data _ in and the label, so that the value of weight is not too large; the value of a may be the average value of data _ in, the value of c may be the average value of label, or the values of a and c may be adjusted according to experience, and will not be described herein again. The smaller b is, the larger alpha is, the larger the weight ratio of the noise-containing part is, the noise suppression is enhanced, and the relative reservation of the target voice is weakened; the smaller d, the larger beta, the larger the weight ratio of the target speech-containing portion, the more speech retention is enhanced, and the relative noise suppression is weakened. Therefore, the aim of controlling the voice enhancement effect can be achieved by adjusting the weight parameters, and each group of weight parameters corresponds to a target voice enhancement effect.
Returning to fig. 1, the training method of the speech enhancement model further includes, in step S106, calculating a second loss value according to the weight value and the first loss value;
optionally, the step S106 includes:
calculating a Hadamard product of the weight value and the first loss value as a second loss value.
It can be understood that, since data _ in, label and the first loss value are vectors, such as one-dimensional vectors or multi-dimensional vectors, a hadamard product of a weight value and the first loss value is calculated as a second loss value, wherein the hadamard product represents a vector obtained by multiplying corresponding positions of two vectors, such as a weight value of [ a ]ij]The matrix is expressed with a first loss value of bij]A matrix of representations, the Hadamard product of the weight value and the first loss value is then given by [ aij*bij]A matrix of representations. The second loss value is added with the weighted value, so that the target voice can be reserved and the noise can be suppressed, and the voice enhancement model capable of realizing the target effect can be obtained by designing the weighted value according to the target effect. For example, to increase the retention of the target speech, a weight value obtained by calculating a weight value according to the weight parameter and the labeled time-frequency mask, that is, a portion of data _ in is omitted, and only a portion of label is retained, where if a is 1, c is 0, d is 1/2, and β is 2:
since label is labeled time-frequency masking and represents the ratio of pure voice, namely the voice spectrum of the target voice to the voice with noise, 0 is larger than or equal to label and smaller than or equal to 1, the larger the target voice proportion is, the larger the value of label is, the larger the value of weight obtained thereby is, and the proportion is amplified through the values of d and beta, so that the larger the energy of the position containing the target voice is in the obtained weight value, the larger the weight value is, and finally the hadamard product is performed on the weight value and the first loss value, so that the place with the target voice has higher weight, and in the convergence process of the model, if the final first loss value is to be as small as possible, the place with the target voice must not be damaged too much, so that the purpose of keeping the target voice is achieved.
Returning to fig. 1, the training method of the speech enhancement model further includes, in step S107, updating the model parameters of the speech enhancement model according to the second loss value;
in this step, the model parameters are updated according to the second loss value through back propagation, and a common back algorithm is gradient descent and the like.
Returning to fig. 1, the training method of the speech enhancement model further includes step S108, iterating the above process from obtaining the prediction time-frequency mask to updating the model parameters until reaching the convergence condition of the speech enhancement model.
In this step, the updated model parameters are used to extract new noisy speech from the training set again, obtain the features of the noisy speech, input into the speech enhancement model, and repeat the updating process of the model parameters, i.e., step S103 to step S107, iteratively update the model parameters until a preset convergence condition is reached, where the convergence condition includes, but is not limited to: the number of iterations exceeds a preset number, the first loss value is less than a preset value, and so on.
Through the steps S101 to S107, the voice enhancement model is obtained through training, and due to the fact that the weighted value designed through the weight parameter is added in the training process to influence the updating of the model parameter, the purpose of controlling the voice enhancement effect of the voice enhancement model is achieved.
Fig. 2 is a logic block diagram of a training method of a speech enhancement model according to an embodiment of the present disclosure. As shown in fig. 2, noisy speech is subjected to feature extraction to obtain input features of a model, the input features are output to the speech enhancement model to obtain model output, the labeled time-frequency mask of the noisy speech and the output of the model are calculated through a loss function to obtain a first loss value loss1, a weight value weight is calculated according to a preset weight parameter, the input features and the labeled time-frequency mask, a product of the weight value and loss1 is calculated to obtain a second loss value loss2, parameters of the model are updated through loss2 and a back propagation algorithm, and the above processes are iterated until loss1 is smaller than a preset value, so that the speech enhancement model with the target enhancement effect is obtained.
The embodiment of the disclosure discloses a training method of a speech enhancement model, which comprises the following steps: acquiring a training data set; acquiring a weight parameter; inputting the characteristics of the voice with noise into a voice enhancement model to obtain prediction time-frequency masking; calculating a first loss value; calculating a weight value according to the weight parameter; calculating a second loss value according to the weight value and the first loss value; updating the model parameters of the voice enhancement model according to the second loss value; and iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached. According to the method, the updating process of the model parameters is influenced through the weight coefficients, so that the voice enhancement model achieves the preset effect, and the technical problem that the voice enhancement effect cannot be controlled when the voice enhancement model is trained in the prior art is solved.
The embodiment of the disclosure also provides a control method of the voice enhancement effect. As shown in fig. 3, the method for controlling the speech enhancement effect includes:
step S301, receiving an input signal of a weight parameter to obtain the weight parameter, wherein the weight parameter corresponds to a target voice enhancement effect;
in this step, a weighting parameter is obtained through an input signal of the human-machine interface, and the weighting parameter corresponds to a specific target voice enhancement effect, such as retaining a target voice or suppressing noise in an enhanced voice.
Step S302, obtaining a voice enhancement model by using the training method of the voice enhancement model according to the weight parameters;
step S303, inputting the voice to be enhanced into the voice enhancement model to obtain prediction time-frequency masking;
and step S304, obtaining the enhanced voice according to the voice to be enhanced and the prediction time-frequency masking.
Step S303 and step S304 are processes of obtaining an enhanced speech by using the speech enhancement model, where the enhanced speech is a target speech, and is a pure de-noised speech that needs to be obtained, and since the predicted time-frequency mask obtained by the speech enhancement model represents a ratio of the target speech in the noisy speech, the pure speech can be obtained by calculating a product of the two.
It can be understood that, the above-mentioned training method of the speech enhancement model can be performed in advance by using multiple sets of weight parameters, resulting in a plurality of speech enhancement models with different effects. Thus, after receiving the input signal of the weighting parameter, the speech enhancement model corresponding to the weighting parameter can be directly selected to realize the control of the sound enhancement effect, which is not described herein again.
In the above, although the steps in the above method embodiments are described in the above sequence, it should be clear to those skilled in the art that the steps in the embodiments of the present disclosure are not necessarily performed in the above sequence, and may also be performed in other sequences such as reverse, parallel, and cross, and further, on the basis of the above steps, other steps may also be added by those skilled in the art, and these obvious modifications or equivalents should also be included in the protection scope of the present disclosure, and are not described herein again.
Fig. 4 is a schematic structural diagram of an embodiment of a training apparatus for a speech enhancement model provided in an embodiment of the present disclosure, and as shown in fig. 4, the apparatus 400 includes: a training set acquisition module 401, a weight parameter acquisition module 402, a prediction module 403, a first loss calculation module 404, a weight calculation module 405, a second loss calculation module 406, a parameter update module 407, and an iteration module 408. Wherein the content of the first and second substances,
a training set obtaining module 401, configured to obtain a training data set, where the training data set includes a noisy speech and a labeled time-frequency mask corresponding to the noisy speech;
a weight parameter obtaining module 402, configured to obtain a weight parameter, where the weight parameter is used to adjust an effect of the speech enhancement model;
a prediction module 403, configured to input the characteristics of the noisy speech into a speech enhancement model to obtain a prediction time-frequency mask;
a first loss calculation module 404, configured to calculate a first loss value according to the labeling time-frequency masking and the prediction time-frequency masking;
a weight calculating module 405, configured to calculate a weight value according to the weight parameter;
a second loss calculating module 406, configured to calculate a second loss value according to the weight value and the first loss value;
a parameter updating module 407, configured to update a model parameter of the speech enhancement model according to the second loss value;
and an iteration module 408, configured to iterate the above process from obtaining the prediction time-frequency mask to updating the model parameters until reaching a convergence condition of the speech enhancement model.
Further, the weight calculating module 405 is further configured to:
calculating a weight value according to the weight parameter and the labeling time-frequency masking; alternatively, the first and second electrodes may be,
calculating a weight value according to the weight parameter and the characteristics of the voice with noise; alternatively, the first and second electrodes may be,
and calculating the weight value according to the weight parameter, the characteristics of the voice with noise and the marking time-frequency mask.
Further, the weight calculating module 405 is further configured to:
according to
Calculating a weight value;
wherein weight represents the weight value, and label represents the labeling time-frequency masking; a, c, d and beta represent the weight parameters, A represents the gain of the weight values, and c, d and beta are used to translate and scale label, where A, c, d and beta are positive numbers.
Further, the weight calculating module 405 is further configured to:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; a, a, b and alpha represent the weight parameters, A represents the gain of the weight values, and a, b and alpha are used for translating and scaling the data _ in, wherein A, a, b and alpha are positive numbers.
Further, the weight calculating module 405 is further configured to:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; label represents the annotation time-frequency masking; a, a, b, c, d, alpha and beta represent the weight parameters, A represents the gain of the weight values, a, b and alpha are used for translating and scaling data _ in, and c, d and beta are used for translating and scaling label, wherein A, a, b, c, d, alpha and beta are positive numbers.
Further, the second loss calculating module 406 is further configured to:
calculating a Hadamard product of the weight value and the first loss value as a second loss value.
The apparatus shown in fig. 4 can perform the method of the embodiment shown in fig. 1 and 2, and the detailed description of this embodiment can refer to the related description of the embodiment shown in fig. 1 and 2. The implementation process and technical effect of the technical solution refer to the description in the embodiment shown in fig. 1 and fig. 2, and are not described herein again.
Fig. 5 is a schematic structural diagram of an embodiment of a control apparatus for a speech enhancement effect provided in an embodiment of the present disclosure, and as shown in fig. 5, the apparatus 500 includes: a weight parameter receiving module 501, a training module 502, an input module 503, and an enhancement module 504. Wherein the content of the first and second substances,
a weight parameter receiving module 501, configured to receive an input signal of a weight parameter to obtain the weight parameter, where the weight parameter corresponds to a target speech enhancement effect;
a training module 502, configured to obtain a speech enhancement model according to the weight parameter by using the training method of the speech enhancement model in the foregoing embodiment;
an input module 503, configured to input a speech to be enhanced into the speech enhancement model to obtain a prediction time-frequency mask;
and an enhancement module 504, configured to obtain an enhanced speech according to the speech to be enhanced and the predicted time-frequency mask.
The apparatus shown in fig. 5 can perform the method of the embodiment shown in fig. 3, and reference may be made to the related description of the embodiment shown in fig. 3 for a part of this embodiment that is not described in detail. The implementation process and technical effect of the technical solution refer to the description in the embodiment shown in fig. 3, and are not described herein again.
Referring now to FIG. 6, a block diagram of an electronic device 600 suitable for use in implementing embodiments of the present disclosure is shown. The terminal device in the embodiments of the present disclosure may include, but is not limited to, a mobile terminal such as a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (personal digital assistant), a PAD (tablet computer), a PMP (portable multimedia player), a vehicle terminal (e.g., a car navigation terminal), and the like, and a stationary terminal such as a digital TV, a desktop computer, and the like. The electronic device shown in fig. 6 is only an example, and should not bring any limitation to the functions and the scope of use of the embodiments of the present disclosure.
As shown in fig. 6, the electronic device may include a processing means (e.g., a central processing unit, a graphics processor, etc.) 601, which may perform various appropriate actions and processes according to a program stored in a Read Only Memory (ROM)602 or a program loaded from a storage means 608 into a Random Access Memory (RAM) 603. In the RAM 603, various programs and data necessary for the operation of the electronic apparatus 600 are also stored. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input/output (I/O) interface 605 is also connected to bus 604.
Generally, the following devices may be connected to the I/O interface 605: input devices 606 including, for example, a touch screen, touch pad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 607 including, for example, a Liquid Crystal Display (LCD), a speaker, a vibrator, and the like; storage 608 including, for example, tape, hard disk, etc.; and a communication device 609. The communication means 609 may allow the electronic device 600 to communicate with other devices wirelessly or by wire to exchange data. While fig. 6 illustrates an electronic device 600 having various means, it is to be understood that not all illustrated means are required to be implemented or provided. More or fewer devices may alternatively be implemented or provided.
In particular, according to an embodiment of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a non-transitory computer readable medium, the computer program containing program code for performing the method illustrated by the flow chart. In such an embodiment, the computer program may be downloaded and installed from a network via the communication means 609, or may be installed from the storage means 608, or may be installed from the ROM 602. The computer program, when executed by the processing device 601, performs the above-described functions defined in the methods of the embodiments of the present disclosure.
It should be noted that the computer readable medium in the present disclosure can be a computer readable signal medium or a computer readable storage medium or any combination of the two. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples of the computer readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present disclosure, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In contrast, in the present disclosure, a computer readable signal medium may comprise a propagated data signal with computer readable program code embodied therein, either in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: electrical wires, optical cables, RF (radio frequency), etc., or any suitable combination of the foregoing.
In some embodiments, the clients, servers may communicate using any currently known or future developed network Protocol, such as HTTP (HyperText Transfer Protocol), and may interconnect with any form or medium of digital data communication (e.g., a communications network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), the Internet (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future developed network.
The computer readable medium may be embodied in the electronic device; or may exist separately without being assembled into the electronic device.
The computer readable medium carries one or more programs which, when executed by the electronic device, cause the electronic device to: the above-described training method of the speech enhancement model is performed.
Computer program code for carrying out operations for the present disclosure may be written in any combination of one or more programming languages, including but not limited to an object oriented programming language such as Java, Smalltalk, C + +, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).
The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The units described in the embodiments of the present disclosure may be implemented by software or hardware. Where the name of an element does not in some cases constitute a limitation on the element itself.
The functions described herein above may be performed, at least in part, by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), systems on a chip (SOCs), Complex Programmable Logic Devices (CPLDs), and the like.
In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
The foregoing description is only exemplary of the preferred embodiments of the disclosure and is illustrative of the principles of the technology employed. It will be appreciated by those skilled in the art that the scope of the disclosure herein is not limited to the particular combination of features described above, but also encompasses other embodiments in which any combination of the features described above or their equivalents does not depart from the spirit of the disclosure. For example, the above features and (but not limited to) the features disclosed in this disclosure having similar functions are replaced with each other to form the technical solution.
Claims (11)
1. A method for training a speech enhancement model, comprising:
acquiring a training data set, wherein the training data set comprises noisy voices and corresponding marked time-frequency masks;
obtaining a weight parameter, wherein the weight parameter is used for adjusting the effect of the voice enhancement model;
inputting the characteristics of the voice with noise into a voice enhancement model to obtain prediction time-frequency masking;
calculating a first loss value according to the labeling time-frequency masking and the prediction time-frequency masking;
calculating a weight value according to the weight parameter;
calculating a second loss value according to the weight value and the first loss value;
updating the model parameters of the voice enhancement model according to the second loss value;
and iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached.
2. The method for training a speech enhancement model according to claim 1, wherein said calculating a weight value according to the weight parameter comprises:
calculating a weight value according to the weight parameter and the labeling time-frequency masking; alternatively, the first and second electrodes may be,
calculating a weight value according to the weight parameter and the characteristics of the voice with noise; alternatively, the first and second electrodes may be,
and calculating the weight value according to the weight parameter, the characteristics of the voice with noise and the marking time-frequency mask.
3. The method for training a speech enhancement model according to claim 2, wherein the calculating a weight value according to the weight parameter and the annotation time-frequency mask comprises:
according to
Calculating a weight value;
wherein weight represents the weight value, and label represents the labeling time-frequency masking; a, c, d and beta represent the weight parameters, A represents the gain of the weight values, and c, d and beta are used for translating and scaling label, wherein A, d and beta are positive numbers, and c is an arbitrary number.
4. The method of claim 2, wherein said calculating weight values based on said weight parameters and said noisy speech features comprises:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; a, a, b and alpha represent the weight parameters, A represents the gain of the weight values, and a, b and alpha are used for translating and scaling the data _ in, wherein A, b and alpha are positive numbers, and a is an arbitrary number.
5. The method of claim 2, wherein said calculating weight values based on said weight parameters and said noisy speech features comprises:
according to
Calculating a weight value;
wherein weight represents the weight value, and data _ in represents the characteristic of the noisy speech; label represents the annotation time-frequency masking; a, a, b, c, d, alpha and beta represent the weight parameters, A represents the gain of the weight values, a, b and alpha are used for translating and scaling data _ in, c, d and beta are used for translating and scaling label, wherein A, b, d, alpha and beta are positive numbers, and a and c are arbitrary numbers.
6. The method of claim 1, wherein the calculating a second loss value based on the weight value and the first loss value comprises:
calculating a Hadamard product of the weight value and the first loss value as a second loss value.
7. A method for controlling a speech enhancement effect, comprising:
receiving an input signal of a weight parameter to obtain the weight parameter, wherein the weight parameter corresponds to a target voice enhancement effect;
obtaining a speech enhancement model according to the weight parameters by using the training method of the speech enhancement model according to claim 1;
inputting the voice to be enhanced into the voice enhancement model to obtain prediction time-frequency masking;
and obtaining the enhanced voice according to the voice to be enhanced and the predicted time-frequency masking.
8. An apparatus for training a speech enhancement model, comprising:
the training set acquisition module is used for acquiring a training data set, wherein the training data set comprises noisy voices and corresponding marked time-frequency masks;
the weight parameter acquisition module is used for acquiring weight parameters, and the weight parameters are used for adjusting the effect of the voice enhancement model;
the prediction module is used for inputting the characteristics of the noisy speech into a speech enhancement model to obtain prediction time-frequency masking;
a first loss calculation module, configured to calculate a first loss value according to the labeling time-frequency mask and the prediction time-frequency mask;
the weight calculation module is used for calculating a weight value according to the weight parameter;
a second loss calculation module, configured to calculate a second loss value according to the weight value and the first loss value;
the parameter updating module is used for updating the model parameters of the voice enhancement model according to the second loss value;
and the iteration module is used for iterating the process from the prediction time frequency masking to the model parameter updating until the convergence condition of the voice enhancement model is reached.
9. An apparatus for controlling a speech enhancement effect, comprising:
the system comprises a weight parameter receiving module, a weight parameter processing module and a voice processing module, wherein the weight parameter receiving module is used for receiving an input signal of a weight parameter to obtain the weight parameter, and the weight parameter corresponds to a target voice enhancement effect;
a training module, configured to obtain a speech enhancement model according to the weight parameter by using the training method of the speech enhancement model according to claim 1;
the input module is used for inputting the voice to be enhanced into the voice enhancement model to obtain prediction time-frequency masking;
and the enhancement module is used for obtaining enhanced voice according to the voice to be enhanced and the prediction time-frequency masking.
10. An electronic device, comprising:
a memory for storing computer readable instructions; and
a processor for executing the computer readable instructions such that the processor when executed implements the method of any of claims 1-7.
11. A computer readable storage medium storing computer readable instructions which, when executed by a computer, cause the computer to perform the method of any one of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011549088.4A CN112669870B (en) | 2020-12-24 | 2020-12-24 | Training method and device for voice enhancement model and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011549088.4A CN112669870B (en) | 2020-12-24 | 2020-12-24 | Training method and device for voice enhancement model and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112669870A true CN112669870A (en) | 2021-04-16 |
| CN112669870B CN112669870B (en) | 2024-05-03 |
Family
ID=75408318
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011549088.4A Active CN112669870B (en) | 2020-12-24 | 2020-12-24 | Training method and device for voice enhancement model and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112669870B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113113038A (en) * | 2021-05-19 | 2021-07-13 | 北京字节跳动网络技术有限公司 | Echo cancellation method and device and electronic equipment |
| CN113436643A (en) * | 2021-06-25 | 2021-09-24 | 平安科技（深圳）有限公司 | Method, device, equipment and storage medium for training and applying speech enhancement model |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20090037692A (en) * | 2007-10-12 | 2009-04-16 | 삼성전자주식회사 | Method and apparatus for extracting the target sound signal from the mixed sound |
| US20160111107A1 (en) * | 2014-10-21 | 2016-04-21 | Mitsubishi Electric Research Laboratories, Inc. | Method for Enhancing Noisy Speech using Features from an Automatic Speech Recognition System |
| US20170140772A1 (en) * | 2015-11-18 | 2017-05-18 | Gwangju Institute Of Science And Technology | Method of enhancing speech using variable power budget |
| CN109036460A (en) * | 2018-08-28 | 2018-12-18 | 百度在线网络技术（北京）有限公司 | Method of speech processing and device based on multi-model neural network |
| CN110251152A (en) * | 2018-03-12 | 2019-09-20 | 北京大学 | A kind of osteoacusis formula number auscultation system being automatically separated heart and lung sounds |
| CN110299149A (en) * | 2019-06-18 | 2019-10-01 | 电子科技大学 | A kind of voice enhancement algorithm based on attention mechanism |
| US20190325889A1 (en) * | 2018-04-23 | 2019-10-24 | Baidu Online Network Technology (Beijing) Co., Ltd | Method and apparatus for enhancing speech |
| CN110580910A (en) * | 2018-06-08 | 2019-12-17 | 北京搜狗科技发展有限公司 | Audio processing method, device and equipment and readable storage medium |
| US20200066296A1 (en) * | 2018-08-21 | 2020-02-27 | 2Hz, Inc | Speech Enhancement And Noise Suppression Systems And Methods |
| CN111192598A (en) * | 2020-01-07 | 2020-05-22 | 哈尔滨理工大学 | Voice enhancement method for jump connection deep neural network |
| CN111429931A (en) * | 2020-03-26 | 2020-07-17 | 云知声智能科技股份有限公司 | Noise reduction model compression method and device based on data enhancement |
| CN111599373A (en) * | 2020-04-07 | 2020-08-28 | 云知声智能科技股份有限公司 | Compression method of noise reduction model |
| CN111613243A (en) * | 2020-04-26 | 2020-09-01 | 云知声智能科技股份有限公司 | Voice detection method and device |
-
2020
- 2020-12-24 CN CN202011549088.4A patent/CN112669870B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20090037692A (en) * | 2007-10-12 | 2009-04-16 | 삼성전자주식회사 | Method and apparatus for extracting the target sound signal from the mixed sound |
| US20160111107A1 (en) * | 2014-10-21 | 2016-04-21 | Mitsubishi Electric Research Laboratories, Inc. | Method for Enhancing Noisy Speech using Features from an Automatic Speech Recognition System |
| US20170140772A1 (en) * | 2015-11-18 | 2017-05-18 | Gwangju Institute Of Science And Technology | Method of enhancing speech using variable power budget |
| CN110251152A (en) * | 2018-03-12 | 2019-09-20 | 北京大学 | A kind of osteoacusis formula number auscultation system being automatically separated heart and lung sounds |
| US20190325889A1 (en) * | 2018-04-23 | 2019-10-24 | Baidu Online Network Technology (Beijing) Co., Ltd | Method and apparatus for enhancing speech |
| CN110580910A (en) * | 2018-06-08 | 2019-12-17 | 北京搜狗科技发展有限公司 | Audio processing method, device and equipment and readable storage medium |
| US20200066296A1 (en) * | 2018-08-21 | 2020-02-27 | 2Hz, Inc | Speech Enhancement And Noise Suppression Systems And Methods |
| CN109036460A (en) * | 2018-08-28 | 2018-12-18 | 百度在线网络技术（北京）有限公司 | Method of speech processing and device based on multi-model neural network |
| CN110299149A (en) * | 2019-06-18 | 2019-10-01 | 电子科技大学 | A kind of voice enhancement algorithm based on attention mechanism |
| CN111192598A (en) * | 2020-01-07 | 2020-05-22 | 哈尔滨理工大学 | Voice enhancement method for jump connection deep neural network |
| CN111429931A (en) * | 2020-03-26 | 2020-07-17 | 云知声智能科技股份有限公司 | Noise reduction model compression method and device based on data enhancement |
| CN111599373A (en) * | 2020-04-07 | 2020-08-28 | 云知声智能科技股份有限公司 | Compression method of noise reduction model |
| CN111613243A (en) * | 2020-04-26 | 2020-09-01 | 云知声智能科技股份有限公司 | Voice detection method and device |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113113038A (en) * | 2021-05-19 | 2021-07-13 | 北京字节跳动网络技术有限公司 | Echo cancellation method and device and electronic equipment |
| CN113436643A (en) * | 2021-06-25 | 2021-09-24 | 平安科技（深圳）有限公司 | Method, device, equipment and storage medium for training and applying speech enhancement model |
| CN113436643B (en) * | 2021-06-25 | 2024-05-24 | 平安科技（深圳）有限公司 | Training and application method, device and equipment of voice enhancement model and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112669870B (en) | 2024-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110600018B (en) | Voice recognition method and device and neural network training method and device | |
| WO2020155907A1 (en) | Method and apparatus for generating cartoon style conversion model | |
| CN112669870B (en) | Training method and device for voice enhancement model and electronic equipment | |
| CN113192528B (en) | Processing method and device for single-channel enhanced voice and readable storage medium | |
| US20240046955A1 (en) | Voice extraction method and apparatus, and electronic device | |
| CN112270200B (en) | Text information translation method and device, electronic equipment and storage medium | |
| CN112752118B (en) | Video generation method, device, equipment and storage medium | |
| CN114020950A (en) | Training method, device and equipment of image retrieval model and storage medium | |
| CN112767959A (en) | Voice enhancement method, device, equipment and medium | |
| CN111312223B (en) | Training method and device of voice segmentation model and electronic equipment | |
| CN111312224B (en) | Training method and device of voice segmentation model and electronic equipment | |
| CN111653261A (en) | Speech synthesis method, speech synthesis device, readable storage medium and electronic equipment | |
| CN114171043B (en) | Echo determination method, device, equipment and storage medium | |
| CN113593527B (en) | Method and device for generating acoustic features, training voice model and recognizing voice | |
| CN111626044B (en) | Text generation method, text generation device, electronic equipment and computer readable storage medium | |
| CN111581455B (en) | Text generation model generation method and device and electronic equipment | |
| CN111898658B (en) | Image classification method and device and electronic equipment | |
| CN110209851B (en) | Model training method and device, electronic equipment and storage medium | |
| CN113780534A (en) | Network model compression method, image generation method, device, equipment and medium | |
| CN113763976A (en) | Method and device for reducing noise of audio signal, readable medium and electronic equipment | |
| CN111582456A (en) | Method, apparatus, device and medium for generating network model information | |
| CN111709366A (en) | Method, apparatus, electronic device, and medium for generating classification information | |
| CN112634930B (en) | Multichannel sound enhancement method and device and electronic equipment | |
| CN117316160B (en) | Silent speech recognition method, silent speech recognition apparatus, electronic device, and computer-readable medium | |
| CN111860518B (en) | Method, apparatus, device and computer readable medium for segmenting an image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |