Voice signal enhancement method, device, storage medium and electronic device

Abstract

The embodiments of the present application disclose a speech signal enhancement method, device, storage medium and electronic device, which can be applied to the cloud field, artificial intelligence field or speech application field. The above method includes: performing feature extraction processing on the speech signal to obtain speech feature information; inputting the above speech feature information into the target model to extract the gain signal to obtain the gain signal corresponding to the above speech signal; enhancing the above speech signal based on the above gain signal to obtain an enhancement result; wherein the above target model includes at least one speech processing module, and the above speech processing module is composed of three gate loop units, and the above gate loop unit is the only speech signal processing unit used in the above target model. The present application not only reduces the degree of heterogeneity and computing requirements, but also improves the denoising and enhancement effect.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a method, device, storage medium, and electronic device for enhancing a speech signal. Background Art

[0002] With the continuous development of science and technology, it has become a general trend to conduct interpersonal communication, information exchange and intelligent control through voice. Voice can free people's hands and provide more convenience for life. With this demand, voice processing technology has also made great progress, among which voice signal enhancement is an important research content. In many cases, the voice quality is relatively low and contains a lot of noise. Voice signal enhancement technology can enhance the effective information in the voice and reduce the noise. Usually, a good voice enhancement effect can be guaranteed by mixing various voice signal processing units. However, this method has high computational complexity and slow speed, and is difficult to apply to application scenarios with high real-time requirements. Summary of the invention

[0003] In order to solve at least one of the above technical problems, the embodiments of the present application provide a speech signal enhancement method, device, storage medium and electronic device.

[0004] On the one hand, an embodiment of the present application provides a method for enhancing a speech signal, the method comprising:

[0005] Perform feature extraction processing on the speech signal to obtain speech feature information;

[0006] Inputting the speech feature information into a target model to extract a gain signal, thereby obtaining a gain signal corresponding to the speech signal;

[0007] Enhance the speech signal based on the gain signal to obtain an enhanced result;

[0008] The target model includes at least one speech processing module, the speech processing module is composed of three gate recurrence units, and the gate recurrence unit is the only speech signal processing unit used in the target model.

[0009] On the other hand, an embodiment of the present application provides a speech signal enhancement device, the device comprising:

[0010] A feature extraction module is used to perform feature extraction processing on the speech signal to obtain speech feature information;

[0011] A gain acquisition module, used for inputting the speech feature information into a target model to extract a gain signal, and obtaining a gain signal corresponding to the speech signal;

[0012] An enhancement module, used for enhancing the speech signal based on the gain signal to obtain an enhancement result;

[0013] The target model includes at least one speech processing module, the speech processing module is composed of three gate recurrence units, and the gate recurrence unit is the only speech signal processing unit used in the target model.

[0014] On the other hand, an embodiment of the present application provides a computer-readable storage medium, in which at least one instruction or at least one program is stored, and the at least one instruction or at least one program is loaded and executed by a processor to implement the above-mentioned speech signal enhancement method.

[0015] On the other hand, an embodiment of the present application provides an electronic device, comprising at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the above-mentioned speech signal enhancement method by executing the instructions stored in the memory.

[0016] On the other hand, an embodiment of the present application provides a computer program product, including a computer program or instructions, which implement the above-mentioned method for enhancing a speech signal when executed by a processor.

[0017] The embodiment of the present application provides a method for speech signal enhancement, which uses a target model obtained by stacking gated recurrent units (GRU) to output a gain signal corresponding to a speech signal, and amplifies the effective content in the speech signal through the gain signal to suppress noise, thereby achieving a better enhanced denoising effect. In addition, the target model only includes a speech signal processing unit, the gated recurrent unit GRU, and does not include other speech signal processing units, such as CNN or LTSM, which reduces the heterogeneity of the target model, reduces the demand for computing resources required in the speech enhancement process, and improves the enhancement speed of the speech signal, so that the speech signal enhancement method provided by the embodiment of the present application can be applied to real-time speech enhancement scenarios, and it is confirmed by experiments that the unique design method of the embodiment of the present application can achieve a denoising enhancement effect similar to or even better than that of similar algorithms at a low computational complexity, which not only reduces the heterogeneity and computing requirements, but also improves the denoising enhancement effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or related technologies, the following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 This is a schematic diagram of a feasible implementation framework of the speech signal enhancement method provided in the embodiments of this specification;

[0020] Figure 2 It is a flow chart of a method for enhancing a speech signal provided in an embodiment of this specification;

[0021] Figure 3 It is a flowchart of a method for extracting features from a speech signal provided in an embodiment of the present application;

[0022] Figure 4 is a schematic diagram of the target model structure provided in an embodiment of the present application;

[0023] FIG5( a) is a schematic diagram of an original speech signal containing noise provided in an embodiment of the present application;

[0024] FIG5( b ) is a schematic diagram of a speech signal obtained after speech enhancement according to an embodiment of the present application;

[0025] Figure 6 is a block diagram of a speech signal enhancement device provided in an embodiment of the present application;

[0026] Figure 7 It is a schematic diagram of the hardware structure of a device provided in an embodiment of the present application for implementing the method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the embodiments of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0029] In order to make the purpose, technical solution and advantages disclosed in the embodiments of the present application more clearly understood, the embodiments of the present application are further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present application and are not used to limit the embodiments of the present application.

[0030] In the following, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this embodiment, unless otherwise specified, "multiple" means two or more. In order to facilitate the understanding of the above-mentioned technical solutions and the technical effects produced by the embodiments of the present application, the embodiments of the present application first explain the relevant professional terms:

[0031] Cloud technology: refers to a hosting technology that unifies hardware, software, network and other resources within a wide area network or local area network to achieve data computing, storage, processing and sharing. It is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the cloud computing business model application, which can form a resource pool and be used on demand, flexible and convenient. Cloud computing technology will become an important support. The background services of the technical network system require a large amount of computing and storage resources, such as video websites, picture websites and more portal websites. With the high development and application of the Internet industry, in the future, each item may have its own identification mark, and all need to be transmitted to the background system for logical processing. Data of different levels will be processed separately. All kinds of industry data require strong system backing support, which can only be achieved through cloud computing.

[0032] Intelligent Traffic System (ITS), also known as Intelligent Transportation System, effectively and comprehensively applies advanced science and technology (information technology, computer technology, data communication technology, sensor technology, electronic control technology, automatic control theory, operations research, artificial intelligence, etc.) to transportation, service control and vehicle manufacturing, strengthens the connection between vehicles, roads and users, and thus forms a comprehensive transportation system that ensures safety, improves efficiency, improves the environment and saves energy.

[0033] Intelligent Vehicle Infrastructure Cooperative Systems (IVICS), referred to as IVICS, is a development direction of Intelligent Transportation Systems (ITS). IVICS uses advanced wireless communications and new generation Internet technologies to implement all-round dynamic real-time information interaction between vehicles and roads, and conducts active vehicle safety control and road cooperative management based on the collection and integration of dynamic traffic information in all time and space, fully realizing the effective coordination of people, vehicles and roads, ensuring traffic safety, and improving traffic efficiency, thus forming a safe, efficient and environmentally friendly road traffic system.

[0034] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0035] Convolutional Neural Networks (CNN) is a type of feedforward neural network with deep structure and convolutional calculations. It is one of the representative algorithms of deep learning. Convolutional neural networks have the ability to represent learning and can classify input information in a translation-invariant manner according to their hierarchical structure. Therefore, they are also called "translation-invariant artificial neural networks."

[0036] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0037] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0038] Deep learning: The concept of deep learning originates from the study of artificial neural networks. A multi-layer perceptron with multiple hidden layers is a deep learning structure. Deep learning discovers distributed feature representations of data by combining low-level features to form more abstract high-level representations of attribute categories or features.

[0039] Recurrent Neural Network (RNN) is a type of recurrent neural network that takes sequence data as input, recursively in the direction of sequence evolution, and all nodes (recurrent units) are connected in a chain. Recurrent neural networks have memory, parameter sharing, and Turing completeness, and have certain advantages in learning the nonlinear characteristics of sequences. Recurrent neural networks have advantages in natural language processing (NLP), such as speech recognition, language modeling, machine translation, and other fields. They are also used in various time series forecasts.

[0040] Long-Short Term Memory RNN (LSTM). LSTM belongs to RNN and is a special implementation of it. The control process of LSTM is similar to that of RNN. They both process data flowing through cells during forward propagation. The difference is that the structure and operation of cells in LSTM are different. RNN performs well in short-term sequence prediction, but long-term sequence prediction is not ideal, so some people proposed LSTM.

[0041] The Gate Recurrent Unit (GRU) is a variant of LSTM, which is proposed to solve problems such as long-term memory and gradient in back propagation. In many cases, the actual performance of GRU and LSTM is almost the same, but GRU is simpler to calculate and easier to implement. GRU has only two gates, which combines the input gate and forget gate in LSTM into one, called the update gate, which controls the amount of data that can be retained to the current moment in the previous memory information, or determines how much information from the previous time step and the current time step will be passed on to the future; the other gate of GRU is called the reset gate, which controls how much past information is forgotten.

[0042] Critical band: Critical band is a term used in audiology and psychoacoustics. It was first proposed by Harvey Fletcher in the 1940s. The cochlea is the sensor organ of hearing in the inner ear. The critical band refers to the frequency bandwidth of the auditory filter generated by the structure of the cochlea. In general, the critical band is the sound frequency band in which the perception of a single tone will be disturbed by the auditory masking of a second single tone.

[0043] Bark domain: Bark frequency domain band, which represents the psychoacoustic scale of sound. In acoustic research, people use auditory filters to simulate different critical frequency bands. Later, researchers found that the structure of the human ear will roughly resonate at multiple frequency points. Based on this phenomenon, the audio signal can be divided into multiple Bark domains. It can be considered that the Bark domain can more realistically reflect the feeling of the human ear to the signal.

[0044] Ultra-wideband: sampling rate is greater than or equal to 32000Hz, bandwidth is greater than or equal to 16000Hz (bandwidth is 1 / 2 sampling rate).

[0045] Speech enhancement: When the speech signal is interfered with or even drowned by various noises, the technology of extracting useful speech signals from the noise background and suppressing and reducing noise interference is used.

[0046] Discrete Cosine Transform (DCT) is a transform related to Fourier Transform, which is similar to Discrete Fourier Transform (DFT), but only uses real numbers. Discrete Cosine Transform is equivalent to a Discrete Fourier Transform of about twice its length, which is performed on a real even function (because the Fourier transform of a real even function is still a real even function), and in some transformations it is necessary to move the input or output position by half a unit (DCT has 8 standard types, 4 of which are common).

[0047] Bark-Frequency Cepstral Coefficients (BFCC) are parameters used to describe the energy distribution of speech in the frequency domain.

[0048] Gene cycle: When a person pronounces a voiced sound, the airflow passes through the glottis, generating a quasi-periodic pulsed airflow. This airflow excites the vocal tract to produce voiced sound, also known as voiced speech, which carries most of the mechanical wave energy in speech. The wavelength of this mechanical wave is called the fundamental wave, and the corresponding period is called the fundamental pitch period. The estimation of the fundamental pitch period is called fundamental pitch detection. The ultimate goal of fundamental pitch detection is to draw a trajectory curve of the fundamental pitch period change that is completely consistent with the wavelength. If this is not possible, try to find a trajectory curve that matches.

[0049] Mean Square Error: MSE, the mean square error in parameter estimation refers to the expected value of the square of the difference between the parameter estimate and the true value of the parameter.

[0050] Activation function: The purpose of the activation function of a neural network is to make the neural network nonlinear. The activation function is continuous and differentiable. Continuous means that when the input value changes slightly, the output value also changes slightly. Differentiable means that there are derivatives everywhere in the domain of definition. Common activation functions are as follows:

[0051] Sigmoid is a smooth step function. Sigmoid can convert any value into a probability between 0 and 1 for binary classification.

[0052] Tanh, or hyperbolic tangent, is similar to Sigmoid with increased amplitude, converting input values ​​to between -1 and 1. The derivative of Tanh ranges from 0 to 1, which is better than Sigmoid's 0 to 1 / 4, and to some extent, alleviates the problem of gradient vanishing. Tanh's output and input can maintain a nonlinear monotonous rise and fall relationship, which is consistent with the gradient solution of the network, has good fault tolerance, and is bounded.

[0053] ReLU, or Rectified Linear Unit, activates some neurons and increases sparsity. For any input x, when x is less than 0, the output value is 0, and when x is greater than 0, the output value is x.

[0054] In the related art, a good speech enhancement effect can usually be ensured by mixing various speech signal processing units. For example, a speech signal processing unit can be constructed through network structures such as CNN, LTSM, GRU, etc., and the speech signal can be enhanced by mixing these speech signal processing units with different network structures. However, this method has high computational complexity and slow speed, and is difficult to apply to application scenarios with high real-time requirements. Moreover, if the mixing method is simplified and the heterogeneity of the network is reduced, the effect of enhanced noise reduction will also be greatly reduced. The computational complexity of mixing speech signal processing units with different structures is high, so the hardware performance requirements are also relatively high. For some smart terminals, it will cause problems of excessive load and poor real-time performance.

[0055] In view of this, an embodiment of the present application provides a method for speech signal enhancement, which uses a target model obtained by stacking gated recurrent units (GRU) to output a gain signal corresponding to a speech signal, and amplifies the effective content in the speech signal through the gain signal to suppress noise, thereby achieving a better enhanced denoising effect. In addition, the target model only includes a speech signal processing unit, the gated recurrent unit GRU, and does not include other speech signal processing units, such as CNN or LTSM, which reduces the heterogeneity of the target model, reduces the demand for computing resources required in the speech enhancement process, and improves the enhancement speed of the speech signal, so that the speech signal enhancement method provided by the embodiment of the present application can be applied to real-time speech enhancement scenarios, and it is confirmed by experiments that the unique design method of the embodiment of the present application can achieve a denoising enhancement effect similar to or even better than that of similar algorithms at a low computational complexity, which not only reduces the heterogeneity and computing requirements, but also improves the denoising enhancement effect.

[0056] The embodiments of the present application can be applied to public cloud, private cloud or hybrid cloud scenarios. For example, the speech signal, target model, or software running based on the speech signal and its related data in the present application can be stored in the above-mentioned public cloud, private cloud or hybrid cloud. Private cloud is to create cloud infrastructure and software and hardware resources within a firewall so that various departments within an organization or enterprise can share resources in the data center. Public cloud usually refers to a cloud that can be used by users provided by a third-party provider. Public cloud can generally be used through the Internet, which may be free or low-cost. The core attribute of public cloud is shared resource service. There are many instances of this cloud, which can provide services in today's entire open public network. Hybrid cloud combines public cloud and private cloud, and is the main model and development direction of cloud computing in recent years. Private clouds are mainly for enterprise users. For security reasons, enterprises prefer to store data in private clouds, but at the same time hope to obtain computing resources from public clouds. In this case, hybrid clouds are increasingly being adopted, which mix and match public and private clouds to achieve the best results. This personalized solution achieves the goal of both saving money and being secure.

[0057] See also Figure 1 , Figure 1 is a schematic diagram of a feasible implementation framework of the speech signal enhancement method provided in the embodiments of this specification, such as Figure 1 As shown, the implementation framework may include at least a client 10 and a voice signal enhancement processor 20, and the client 10 and the voice signal enhancement processor 20 communicate through a network 30. The voice signal enhancement processor 20 may perform denoising and enhancement processing on the voice signal acquired in the client 10 to obtain an enhanced voice signal. Of course, the voice signal enhancement processor 20 may store the enhanced voice signal locally, transmit it back to the client 10, or transmit it to other clients 10. The above-mentioned framework of the embodiment of the present invention may provide the voice signal enhancement capability required for applications in various scenarios, including but not limited to cloud technology, cloud gaming, cloud rendering, artificial intelligence, smart transportation, assisted driving, video media, smart communities, instant messaging, etc. Each component in the framework may be a terminal device or a server. Terminal devices include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, etc.

[0058] The following describes a method for enhancing a speech signal according to an embodiment of the present application. Figure 2A flow chart of a speech signal enhancement method provided by an embodiment of the present application is shown. The embodiment of the present application provides the method operation steps as described in the embodiment or flowchart, but more or fewer operation steps may be included based on conventional or non-creative labor. The order of steps listed in the embodiment is only one way of executing the order of many steps and does not represent the only execution order. When the actual system, terminal device or server product is executed, it can be executed in sequence or in parallel (for example, in a parallel processor or multi-threaded processing environment) according to the method shown in the embodiment or the accompanying drawings. The above method may include:

[0059] S101. Perform feature extraction processing on the speech signal to obtain speech feature information.

[0060] The embodiment of the present application does not limit the method for extracting speech feature information. For example, the Bark frequency domain band (Bark domain, or Bark domain) can be divided according to the frequency domain range covered by the speech signal, and some parameters can be extracted for each Bark frequency domain band, such as spectrum amplitude value vector, spectrum logarithmic energy value vector, filter bank (Filter Banks, Fbanks) vector, Bark frequency cepstral coefficient (Bark-Frequency Cepstral Coefficients, BFCC) vector, fundamental pitch period, etc. The embodiment of the present application does not limit the Bark domain division method, and reference can be made to related technologies.

[0061] In a specific embodiment, Figure 3 As shown, the method for extracting features from a speech signal to obtain speech feature information specifically includes:

[0062] S1011. Perform frequency domain bandwidth division on the above-mentioned speech signal to obtain a first number of Bark frequency domain bands.

[0063] The embodiment of the present application can divide the Bark frequency domain band according to the frequency domain range covered by the voice signal. The embodiment of the present application does not limit the first number. For example, 24 widely used in the related art can be adopted, or 29 can be used according to the ultra-wide bandwidth scenario used in this application.

[0064] S1012. For each Bark frequency domain band, determine its corresponding Bark frequency cepstral coefficient.

[0065] Taking the first number of 29 as an example, 29 Bark frequency cepstral coefficients BFCC can be obtained.

[0066] S1013. For each of the first second number of Bark frequency domain bands of the frequency domain ascending arrangement result of the first number of Bark frequency domain bands, obtain its corresponding first-order difference parameter, second-order difference parameter, and cross-correlation parameter.

[0067] The embodiment of the present application does not limit the specific value of the second number, and can be set according to actual needs. Taking 29 Bark frequency domain bands as an example, the first 6 Bark frequency domain bands can be taken to calculate their corresponding first-order difference parameters, second-order difference parameters and cross-correlation parameters, where the cross-correlation parameters can be understood as DCT transformation values, which results in 18 parameters.

[0068] S1014. Obtain the gene period and energy parameters of the above speech signal.

[0069] The embodiments of the present application do not limit the method for obtaining the gene period and energy parameters of the speech signal, and reference may be made to related technologies.

[0070] S1015. Construct the above-mentioned speech feature information according to the acquired Bark frequency cepstral coefficients, first-order difference parameters, second-order difference parameters, cross-correlation parameters, gene periodicity and energy parameters.

[0071] This application does not limit the construction method, and the various information obtained can be combined together to construct speech feature information in the form of a matrix or vector. Taking 29 Bark frequency domain bands as an example, the 29 BFCCs obtained in step S1012, the 18 parameters obtained in step S1013, and the two parameters obtained in step S1014 are a total of 49 parameters, which can construct 49-dimensional speech feature information. This method of constructing speech features can make the constructed speech feature information cover comprehensive valid data, which is conducive to improving the accuracy of gain signal acquisition.

[0072] S102. Input the above-mentioned speech feature information into the target model to extract the gain signal to obtain the gain signal corresponding to the above-mentioned speech signal. The above-mentioned target model includes at least one speech processing module, and the above-mentioned speech processing module is composed of three gate loop units, and the above-mentioned gate loop unit is the only speech signal processing unit used in the above-mentioned target model.

[0073] The embodiment of the present application extracts the gain signal based on the target model of the speech signal processing unit with only a single structure, thereby ensuring low computational consumption and high enhancement speed. Specifically, the target model includes at least one speech processing module, the speech processing module is composed of three gated recurrent units, and the gated recurrent unit is the only speech signal processing unit used in the target model.

[0074] The target model may include only one speech processing module or at least two speech processing modules. When the target model includes at least two speech processing modules, the speech processing modules in the target model form a unidirectional chain connection; a connection layer is provided between any two adjacent speech processing modules in the target model, and the connection layer is composed of the gate loop unit.

[0075] Each speech processing module includes a first gate recurrence unit, a second gate recurrence unit and a third gate recurrence unit, and the gate recurrence units with corresponding relationships in different speech processing modules have the same number of nodes.

[0076] by Figure 4 As an example, a schematic diagram of the target model structure is shown. The target model has two speech processing modules, each of which includes three gated recurrent units (GRUs). Two adjacent speech processing modules are connected through a connection layer composed of gated recurrent units. Therefore, the target model includes a total of 7 layers of gated recurrent units, and the purpose of speech signal processing is achieved by stacking gated recurrent units.

[0077] The speech feature information is first input into the first speech processing module, and the output of the first speech processing module is transmitted to the second speech processing module through the connection layer. The output of the second speech processing module can obtain a gain signal after being activated.

[0078] The embodiments of the present application do not limit the activation functions of the gated recurrent units in the speech processing module and the activation function of the final activation layer, but it should be noted that the activation functions of the gated recurrent units with corresponding relationships in different speech processing modules should be the same, and the number of nodes should also be consistent. For example, for each speech processing module, the activation functions corresponding to the first gated recurrent unit, the second gated recurrent unit, and the third gated recurrent unit can be Tanh, ReLU, and ReLU, respectively.

[0079] The embodiment of the present application does not limit the number of nodes of each gate cycle unit. Figure 4 Taking the first speech processing module in as an example, the first gate recurrent unit is GRU tanh(60), indicating that its activation function is Tanh and it has 60 nodes. The second gate recurrent unit is GRU ReLU(80), indicating that its activation function is ReLU and it has 80 nodes. The third gate recurrent unit is GRU ReLU(140), indicating that its activation function is ReLU and it has 140 nodes. The connection layer is GRU tanh(60), indicating that its activation function is Tanh and it has 60 nodes. The output of the target model is Densesigmoid(29), indicating that the activation layer is in a dense parameter state, the activation function is Sigmoid and it has 29 nodes.

[0080] The embodiment of the present application does not limit the number of nodes and activation functions of each gate recurrent unit. For example, the number of nodes in the activation layer can also be 60, and the activation function used by the gate recurrent unit can also be changed, such as Figure 4GRU ReLU / tanh is used in the training, and its activation function can also be changed to Sigmoid or ReLU variants.

[0081] In the present application embodiment Figure 4 The target model shown has a relatively reasonable structure, whether it is the number of nodes in each layer, complexity, vertical reasoning depth, number of nodes, etc., Figure 4 The actual verification results of the target model show that it can achieve similar or better denoising results than similar algorithms.

[0082] In one embodiment, when the target model includes a speech processing module, the speech feature information is input into the target model for gain signal extraction to obtain a gain signal corresponding to the speech signal, including: using the speech feature information as the input signal of the first speech processing module in the target model to obtain a processing result output by the first speech processing module; and inputting the processing result into the activation function output layer of the target model to obtain the gain signal.

[0083] In another embodiment, when the target model includes at least two speech processing modules, the speech feature information is input into the target model for gain signal extraction to obtain a gain signal corresponding to the speech signal, including: using the speech feature information as the input signal of the first speech processing module in the target model to obtain the processing result output by the first speech processing module; inputting the processing result into the connection layer to obtain the input signal of the second speech processing unit; obtaining the processing result output by the last speech processing unit based on the processing result output by the second speech processing unit; and inputting the processing result output by the last speech processing unit into the activation function output layer of the target model to obtain the gain signal.

[0084] by Figure 4 For example, the gain signal can be obtained through the following steps:

[0085] Step 1: Extract the speech and obtain the speech signal. Extract 49-dimensional features from each frame of speech signal and construct speech feature information.

[0086] Step 2: Input the extracted speech feature information into the first gate recurrent unit GRU tanh (60) in the first speech processing module of the target model.

[0087] Step 3: Pass the output of GRU tanh (60) and the speech feature information constructed by 49-dimensional features to the second gate recurrent unit GRU ReLU (80) in the first speech processing module of the target model.

[0088] Step 4: The output of GRU ReLU (80) is combined with the output of the first layer GRU tanh (60) and the speech feature information constructed by 49-dimensional features, and passed to the third gate recurrent unit GRU tanh (140) in the first speech processing module of the target model.

[0089] Step 5: The output of GRU ReLU (140) is then input to the connection layer GRU tanh (60).

[0090] Step 6: The output of the connection layer GRU (60) is directly used as the input information of the second speech processing module of the target model, and is input into the first gate recurrent unit GRU tanh (60) in the second speech processing module of the target model.

[0091] Step 7: The output of GRU (60) and the output of the connection layer (as input information of the second speech processing module) are passed to the second gate recurrent unit GRU ReLU (80) in the second speech processing module of the target model.

[0092] Step 8: The output of GRU tanh (80) is combined with the output of GRU (60) and the output of the connection layer, and passed to the third gate recurrent unit GRU tanh (140) in the second speech processing module of the target model.

[0093] Step 9: The output of the third gate recurrent unit GRU tanh (140) in the second speech processing module of the target model is passed through a Dense Sigmoid activation function to obtain a gain signal.

[0094] Of course, the embodiment of the present application is described by taking the target model including two speech processing modules as an example. If more speech processing modules are included, the same can be applied and will not be elaborated in this article.

[0095] As can be seen from the foregoing, for each speech processing module, the same processing method is adopted according to its input information. Taking the first speech processing module in the target model as an example, the above-mentioned speech feature information is used as the input signal of the first speech processing module in the above-mentioned target model to obtain the processing result output by the above-mentioned first speech processing module, including: inputting the above-mentioned input signal into the above-mentioned first gate loop unit to obtain the first signal; inputting the above-mentioned input signal and the above-mentioned first signal into the above-mentioned second gate loop unit to obtain the second signal; inputting the above-mentioned first signal, the above-mentioned second signal and the above-mentioned input signal into the above-mentioned third gate loop unit to obtain the processing result output by the above-mentioned first speech processing module.

[0096] The embodiments of the present application do not limit the training method of the target model. Specifically, the target model is trained by the following method: obtaining a speech sample signal and a pure sample signal corresponding to the speech sample signal, the pure sample signal being the speech sample signal that does not contain noise; inputting the speech sample signal into a preset model to obtain a sample gain signal, the preset model and the target model have the same structure; obtaining a sample true value gain based on the speech sample signal and the pure sample signal; determining the loss of the preset model based on the sample gain signal and the sample true value gain; adjusting the parameters of the preset model based on the loss to obtain the target model.

[0097] The sample truth gain can be calculated by the pure speech and noisy speech signals, and the embodiment of the present application does not limit the method for determining the loss, for example, it can use the cross entropy, MSE cost function, etc. The embodiment of the present application does not limit the termination condition of the training, for example, the training can be terminated after the number of parameter adjustments reaches a preset number or the loss reaches a preset requirement to obtain the target model.

[0098] S103. Enhance the speech signal based on the gain signal to obtain an enhanced result.

[0099] Specifically, the power spectrum of the speech signal may be obtained; the gain signal and the power spectrum may be multiplied to obtain a frequency domain gain signal; and the frequency domain gain signal may be discretely cosine transformed to obtain the enhancement result.

[0100] Once the target model is trained, it can be used for online enhancement processing. It only needs to extract the corresponding speech signal features from the noisy speech signal and then input the features into the trained target model. For example, Figure 4 The target model can be adaptively adapted to obtain the optimal gain information. Finally, the gain is multiplied by the power spectrum of the noisy speech signal, and the final enhanced speech signal is obtained by inverse FFT transformation.

[0101] FIG5(a) is a schematic diagram of an original speech signal containing noise. FIG5(b) is a schematic diagram of a speech signal obtained after speech enhancement by the embodiment of the present application. Combining FIG5(a) and FIG5(b) can reflect the speech enhancement effect of the embodiment of the present application. In the speech segment, the noise removed by the method of the embodiment of the present application is cleaner. At the same time, if the method of the embodiment of the present application is used Figure 4 The target model of the structure can control the number of model parameters to around 462,000 (462K), which is actually a low-complexity enhancement algorithm.

[0102] The embodiment of the present application provides a method for speech signal enhancement, which uses a target model obtained by stacking gated recurrent units (GRU) to output a gain signal corresponding to a speech signal, and amplifies the effective content in the speech signal through the gain signal to suppress noise, thereby achieving a better enhanced denoising effect. In addition, the target model only includes a speech signal processing unit, the gated recurrent unit GRU, and does not include other speech signal processing units, such as CNN or LTSM, which reduces the heterogeneity of the target model, reduces the demand for computing resources required in the speech enhancement process, and improves the enhancement speed of the speech signal, so that the speech signal enhancement method provided by the embodiment of the present application can be applied to real-time speech enhancement scenarios, and it is confirmed by experiments that the unique design method of the embodiment of the present application can achieve a denoising enhancement effect similar to or even better than that of similar algorithms at a low computational complexity, which not only reduces the heterogeneity and computing requirements, but also improves the denoising enhancement effect.

[0103] The algorithm of the embodiment of the present application has good real-time performance, so it can be applied to more usage scenarios without being constrained by real-time requirements. For example, when the call environment is noisy, the method proposed in the embodiment of the present application can effectively remove background environmental noise, thereby ensuring the quality and intelligibility of voice calls. When the microphone is turned on, the sound signal is collected by the microphone, and then the signal is feature extracted and passed through the target model to obtain a gain signal. Based on the gain signal, a relatively "clean" voice signal can be obtained. The voice signal can be regarded as an enhanced voice signal, which can effectively retain the voice of the main speaker and remove excess background noise.

[0104] Please refer to Figure 6 , which shows a block diagram of a speech signal enhancement device in this embodiment, the device includes:

[0105] The feature extraction module 101 is used to perform feature extraction processing on the speech signal to obtain speech feature information;

[0106] The gain acquisition module 102 is used to input the above-mentioned speech feature information into the target model to extract the gain signal, so as to obtain the gain signal corresponding to the above-mentioned speech signal;

[0107] An enhancement module 103, configured to enhance the speech signal based on the gain signal to obtain an enhancement result;

[0108] The target model includes at least one speech processing module, which is composed of three gate recurrence units. Furthermore, the gate recurrence unit is the only speech signal processing unit used in the target model.

[0109] In one embodiment, the target model includes at least two speech processing modules.

[0110] Each speech processing module in the above target model forms a one-way chain connection;

[0111] A connection layer is provided between any two adjacent speech processing modules in the target model, and the connection layer is composed of the gate recurrent unit.

[0112] In one embodiment, each speech processing module includes a first gate recurrence unit, a second gate recurrence unit and a third gate recurrence unit, and the gate recurrence units with corresponding relationships in different speech processing modules have the same number of nodes.

[0113] In one embodiment, the activation functions corresponding to the first gate recurrent unit, the second gate recurrent unit, and the third gate recurrent unit are Tanh, ReLU, and ReLU, respectively.

[0114] In one embodiment, when the target model includes a speech processing module,

[0115] The gain acquisition module 102 is used to:

[0116] The above-mentioned inputting the above-mentioned speech feature information into the target model to extract the gain signal to obtain the gain signal corresponding to the above-mentioned speech signal includes:

[0117] Using the speech feature information as an input signal of a first speech processing module in the target model to obtain a processing result output by the first speech processing module;

[0118] The above processing results are input into the activation function output layer of the above target model to obtain the above gain signal.

[0119] In one embodiment, when the target model includes at least two speech processing modules, the gain acquisition module 102 is used to:

[0120] Using the speech feature information as an input signal of a first speech processing module in the target model to obtain a processing result output by the first speech processing module;

[0121] The above processing results are input into the above connection layer to obtain the input signal of the second speech processing unit:

[0122] Based on the processing result output by the second speech processing unit, obtaining the processing result output by the last speech processing unit;

[0123] The processing result output by the last speech processing unit is input into the activation function output layer of the target model to obtain the gain signal.

[0124] In one embodiment, the gain acquisition module 102 is used to:

[0125] Input the input signal into the first gate cycle unit to obtain a first signal;

[0126] Input the input signal and the first signal into the second gate cycle unit to obtain a second signal;

[0127] The first signal, the second signal and the input signal are all input into the third gate loop unit to obtain the processing result output by the first speech processing module.

[0128] In one embodiment, the feature extraction module 101 is used to:

[0129] Performing frequency domain bandwidth division on the speech signal to obtain a first number of Bark frequency domain bands;

[0130] For each Bark frequency domain band, determine the corresponding Bark frequency cepstral coefficient;

[0131] For each of the first second number of Bark frequency domain bands of the frequency domain ascending order arrangement result of the first number of Bark frequency domain bands, obtain the corresponding first-order difference parameter, second-order difference parameter, and cross-correlation parameter;

[0132] Obtaining gene period and energy parameters of the above speech signal;

[0133] The above-mentioned speech feature information is constructed according to the obtained Bark frequency cepstral coefficients, first-order difference parameters, second-order difference parameters, cross-correlation parameters, gene period and energy parameters.

[0134] In one embodiment, the enhancement module 103 is used to:

[0135] Obtaining the power spectrum of the speech signal;

[0136] Multiplying the gain signal by the power spectrum to obtain a frequency domain gain signal;

[0137] The frequency domain gain signal is subjected to discrete cosine transform to obtain the enhancement result.

[0138] In one embodiment, the target model is trained by the following method:

[0139] Acquire a speech sample signal and a clean sample signal corresponding to the speech sample signal, wherein the clean sample signal is the speech sample signal without noise;

[0140] Inputting the speech sample signal into a preset model to obtain a sample gain signal, wherein the preset model has the same structure as the target model;

[0141] According to the above-mentioned speech sample signal and the above-mentioned pure sample signal, a sample true value gain is obtained;

[0142] Determining the loss of the preset model according to the sample gain signal and the sample true value gain;

[0143] The parameters of the preset model are adjusted according to the loss to obtain the target model. The device part of the embodiment of the present application is based on the same inventive concept as the method embodiment, and will not be described in detail here.

[0144] Further, Figure 7 A schematic diagram of the hardware structure of a device for implementing the method provided in the embodiment of the present application is shown. The above device may participate in constituting or include the apparatus or system provided in the embodiment of the present application. Figure 7 As shown, the device 10 may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 7 The structure shown is for illustration only and does not limit the structure of the electronic device. Figure 7 More or fewer components as shown, or with Figure 7 Different configurations are shown.

[0145] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the device 10 (or mobile device). As involved in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0146] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the above-mentioned method in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, to implement the above-mentioned method for enhancing a voice signal. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the device 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0147] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the device 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet wirelessly.

[0148] The display may be, for example, a touch screen liquid crystal display (LCD) that may enable a user to interact with a user interface of device 10 (or mobile device).

[0149] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of the embodiments of the present application are described. Other embodiments are within the scope of the attached claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0150] Each embodiment in the embodiments of the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and server embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0151] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program. The above program may be stored in a computer-readable storage medium, and the above storage medium may be a read-only memory, a disk or an optical disk, etc.

[0152] The instructions in the above storage medium can execute a method for enhancing a speech signal, the method comprising:

[0153] Perform feature extraction processing on the speech signal to obtain speech feature information;

[0154] Inputting the above speech feature information into the target model to extract the gain signal, thereby obtaining a gain signal corresponding to the above speech signal;

[0155] The speech signal is enhanced based on the gain signal to obtain an enhanced result;

[0156] The target model includes at least one speech processing module, which is composed of three gate recurrence units. Furthermore, the gate recurrence unit is the only speech signal processing unit used in the target model.

[0157] In one embodiment, the target model includes at least two speech processing modules.

[0158] Each speech processing module in the above target model forms a one-way chain connection;

[0159] A connection layer is provided between any two adjacent speech processing modules in the target model, and the connection layer is composed of the gate recurrent unit.

[0160] In one embodiment, each speech processing module includes a first gate recurrence unit, a second gate recurrence unit and a third gate recurrence unit, and the gate recurrence units with corresponding relationships in different speech processing modules have the same number of nodes.

[0161] In one embodiment, the activation functions corresponding to the first gate recurrent unit, the second gate recurrent unit, and the third gate recurrent unit are Tanh, ReLU, and ReLU, respectively.

[0162] In one embodiment, when the target model includes a speech processing module,

[0163] The above-mentioned inputting the above-mentioned speech feature information into the target model to extract the gain signal to obtain the gain signal corresponding to the above-mentioned speech signal includes:

[0164] Using the speech feature information as an input signal of a first speech processing module in the target model to obtain a processing result output by the first speech processing module;

[0165] The above processing results are input into the activation function output layer of the above target model to obtain the above gain signal.

[0166] In one embodiment, when the target model includes at least two speech processing modules,

[0167] The above-mentioned inputting the above-mentioned speech feature information into the target model to extract the gain signal to obtain the gain signal corresponding to the above-mentioned speech signal includes:

[0168] Using the speech feature information as an input signal of a first speech processing module in the target model to obtain a processing result output by the first speech processing module;

[0169] The above processing results are input into the above connection layer to obtain the input signal of the second speech processing unit:

[0170] Based on the processing result output by the second speech processing unit, obtaining the processing result output by the last speech processing unit;

[0171] The processing result output by the last speech processing unit is input into the activation function output layer of the target model to obtain the gain signal.

[0172] In one embodiment, the above-mentioned speech feature information is used as an input signal of the first speech processing module in the above-mentioned target model to obtain the processing result output by the above-mentioned first speech processing module, including:

[0173] Input the input signal into the first gate cycle unit to obtain a first signal;

[0174] Input the input signal and the first signal into the second gate cycle unit to obtain a second signal;

[0175] The first signal, the second signal and the input signal are all input into the third gate loop unit to obtain the processing result output by the first speech processing module.

[0176] In one embodiment, the feature extraction process of the speech signal to obtain speech feature information includes:

[0177] Performing frequency domain bandwidth division on the speech signal to obtain a first number of Bark frequency domain bands;

[0178] For each Bark frequency domain band, determine the corresponding Bark frequency cepstral coefficient;

[0179] For each of the first second number of Bark frequency domain bands of the frequency domain ascending order arrangement result of the first number of Bark frequency domain bands, obtain the corresponding first-order difference parameter, second-order difference parameter, and cross-correlation parameter;

[0180] Obtaining gene period and energy parameters of the above speech signal;

[0181] The above-mentioned speech feature information is constructed according to the obtained Bark frequency cepstral coefficients, first-order difference parameters, second-order difference parameters, cross-correlation parameters, gene period and energy parameters.

[0182] In one embodiment, the speech signal is enhanced based on the gain signal to obtain an enhancement result, including:

[0183] Obtaining the power spectrum of the speech signal;

[0184] Multiplying the gain signal by the power spectrum to obtain a frequency domain gain signal;

[0185] The frequency domain gain signal is subjected to discrete cosine transform to obtain the enhancement result.

[0186] In one embodiment, the target model is trained by the following method:

[0187] Acquire a speech sample signal and a clean sample signal corresponding to the speech sample signal, wherein the clean sample signal is the speech sample signal without noise;

[0188] Inputting the speech sample signal into a preset model to obtain a sample gain signal, wherein the preset model has the same structure as the target model;

[0189] According to the above-mentioned speech sample signal and the above-mentioned pure sample signal, a sample true value gain is obtained;

[0190] Determining the loss of the preset model according to the sample gain signal and the sample true value gain;

[0191] The parameters of the preset model are adjusted according to the above loss to obtain the above target model.

[0192] The above is only a preferred embodiment of the embodiment of the present application and is not intended to limit the embodiment of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiment of the present application should be included in the protection scope of the embodiment of the present application.

Claims

1. A speech signal enhancement method, characterized in that: The method comprises: Perform feature extraction processing on the speech signal to obtain speech feature information; Inputting the speech feature information into a target model to extract a gain signal, thereby obtaining a gain signal corresponding to the speech signal; Enhance the speech signal based on the gain signal to obtain an enhanced result; Among them, the target model includes at least two speech processing modules, and the speech processing modules in the target model form a unidirectional chain connection; a connection layer is arranged between any two adjacent speech processing modules in the target model, and the connection layer is composed of a gate loop unit; the speech processing module is composed of three gate loop units, and the gate loop unit is the only speech signal processing unit used in the target model.

2. The method according to claim 1, characterized in that: Each speech processing module includes a first gate recurrence unit, a second gate recurrence unit and a third gate recurrence unit, and the gate recurrence units with corresponding relationships in different speech processing modules have the same number of nodes.

3. The method according to claim 2, characterized in that The activation functions corresponding to the first gate recurrent unit, the second gate recurrent unit, and the third gate recurrent unit are Tanh, ReLU, and ReLU, respectively.

4. The method according to claim 3, characterized in that The step of inputting the speech feature information into a target model to extract a gain signal to obtain a gain signal corresponding to the speech signal comprises: Using the speech feature information as an input signal of a first speech processing module in the target model to obtain a processing result output by the first speech processing module; The processing result is input into the connection layer to obtain the input signal of the second speech processing unit: Based on the processing result output by the second speech processing unit, obtaining the processing result output by the last speech processing unit; The processing result output by the last speech processing unit is input into the activation function output layer of the target model to obtain the gain signal.

5. The method according to claim 4, characterized in that The step of using the speech feature information as an input signal of a first speech processing module in the target model to obtain a processing result output by the first speech processing module includes: Inputting the input signal into the first gate cycle unit to obtain a first signal; Inputting the input signal and the first signal into the second gate cycle unit to obtain a second signal; The first signal, the second signal and the input signal are all input into the third gate loop unit to obtain the processing result output by the first speech processing module.

6. The method according to claim 1, characterized in that The feature extraction process of the speech signal to obtain speech feature information includes: Performing frequency domain bandwidth division on the speech signal to obtain a first number of Bark frequency domain bands; For each Bark frequency domain band, determine the corresponding Bark frequency cepstral coefficient; For each of the first second number of Bark frequency domain bands of the result of the frequency domain ascending arrangement of the first number of Bark frequency domain bands, obtaining a first-order difference parameter, a second-order difference parameter, and a cross-correlation parameter corresponding thereto; Acquiring gene period and energy parameters of the speech signal; The speech feature information is constructed according to the acquired Bark frequency cepstral coefficients, first-order difference parameters, second-order difference parameters, cross-correlation parameters, gene periodicity and energy parameters.

7. The method according to claim 1, characterized in that The step of enhancing the speech signal based on the gain signal to obtain an enhancement result includes: Acquire a power spectrum of the speech signal; Multiplying the gain signal by the power spectrum to obtain a frequency domain gain signal; Performing discrete cosine transform on the frequency domain gain signal to obtain the enhancement result.

8. The method according to claim 1, characterized in that: The target model is trained by the following method: Acquire a speech sample signal and a clean sample signal corresponding to the speech sample signal, wherein the clean sample signal is the speech sample signal without noise; Inputting the speech sample signal into a preset model to obtain a sample gain signal, wherein the preset model has the same structure as the target model; Obtaining a sample true value gain according to the speech sample signal and the clean sample signal; Determining the loss of the preset model according to the sample gain signal and the sample true value gain; The parameters of the preset model are adjusted according to the loss to obtain the target model.

9. A speech signal enhancement device, characterized in that: The device comprises: A feature extraction module is used to perform feature extraction processing on the speech signal to obtain speech feature information; A gain acquisition module, used for inputting the speech feature information into a target model to extract a gain signal, and obtaining a gain signal corresponding to the speech signal; An enhancement module, used for enhancing the speech signal based on the gain signal to obtain an enhancement result; Among them, the target model includes at least two speech processing modules, and the speech processing modules in the target model form a unidirectional chain connection; a connection layer is arranged between any two adjacent speech processing modules in the target model, and the connection layer is composed of a gate loop unit; the speech processing module is composed of three gate loop units, and the gate loop unit is the only speech signal processing unit used in the target model.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by a processor to implement a speech signal enhancement method as described in any one of claims 1 to 8.

11. An electronic device, characterized in that: The invention comprises at least one processor and a memory which is communicatively connected to the at least one processor; wherein the memory stores instructions which can be executed by the at least one processor, and the at least one processor implements a speech signal enhancement method as described in any one of claims 1 to 8 by executing the instructions stored in the memory.

12. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, a speech signal enhancement method according to any one of claims 1 to 8 is implemented.

Images