Voice signal processing method and apparatus, electronic device, and non-volatile storage medium

By extracting acoustic features of speech signals using an unsupervised pre-trained model and combining them with speech quality assessment and sound source distance estimation models, background human voice interference is filtered out, solving the interference problem of speech human-computer interaction systems in complex acoustic scenarios and improving user experience.

WO2026232736A1PCT designated stage Publication Date: 2026-11-12CHINA TELECOM ARTIFICIAL INTELLIGENCE TECHNOLOGY (BEIJING) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/122588
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-07
Filing Date
2025-09-19
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

In complex acoustic scenarios, voice-based human-computer interaction systems are easily affected by background noise, especially since background human voices are highly similar to effective human voices in terms of audio features, making them difficult to be effectively filtered by voice activity detection models in related technologies, resulting in a poor user voice interaction experience.

Method used

An unsupervised pre-trained model is used to extract acoustic features of the speech signal. The speech clarity and source distance are jointly estimated by the speech quality assessment model and the source distance estimation model to filter out interfering speech streams and obtain the target speech signal.

Benefits of technology

It effectively processes speech signals in complex acoustic scenarios, improves the performance of voice human-computer interaction systems, and enhances user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025122588_12112026_PF_FP_ABST
    Figure CN2025122588_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A voice signal processing method and apparatus, an electronic device, and a non-volatile storage medium. The method comprises: acquiring an original voice signal, and performing feature extraction on the original voice signal to obtain acoustic features corresponding to the original voice signal (S202); analyzing the acoustic features by using a voice signal processing model, to obtain a first parameter and a second parameter corresponding to the original voice signal, wherein the first parameter is used for characterizing voice quality of a voice stream in the original voice signal, and the second parameter is used for characterizing a sound source distance corresponding to the voice stream in the original voice signal (S204); and determining an interfering voice stream in the original voice signal on the basis of the first parameter and the second parameter, and filtering out the interfering voice stream from the original voice signal to obtain a target voice signal (S206).
Need to check novelty before this filing date? Find Prior Art

Description

Speech signal processing methods, devices, electronic equipment, and non-volatile storage media

[0001] Related applications

[0002] This application claims priority to Chinese patent application filed on May 7, 2025, with application number 202510586177.2, entitled "Speech Signal Processing Method, Apparatus, Electronic Device and Non-Volatile Storage Medium", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of data processing technology, and more specifically, to a speech signal processing method, apparatus, electronic device, and non-volatile storage medium. Background Technology

[0004] In recent years, breakthroughs in artificial intelligence and deep learning technologies have significantly improved the performance of voice-based human-computer interaction systems. However, in complex acoustic environments, voice-based human-computer interaction is often susceptible to interference from background noise. Furthermore, due to the high similarity in audio features between background voices and valid voices, they are difficult to filter out by voice activity detection (VAD) models in related technologies, resulting in a poor user experience.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This application provides a speech signal processing method, apparatus, electronic device, and non-volatile storage medium.

[0007] According to one aspect of the embodiments of this application, a speech signal processing method is provided, comprising: acquiring an original speech signal and extracting features from the original speech signal to obtain acoustic features corresponding to the original speech signal; using a speech signal processing model to analyze the acoustic features to obtain a first parameter and a second parameter corresponding to the original speech signal, wherein the first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal, the sound source distance being the distance between the sound source generating the speech stream and the sound pickup device acquiring the speech stream; determining the interfering speech stream in the original speech signal based on the first parameter and the second parameter, and filtering out the interfering speech stream from the original speech signal to obtain a target speech signal.

[0008] In some embodiments, the speech signal processing model includes: a pre-trained model, a speech quality assessment model, and a source distance estimation model; the training steps of the speech signal processing model include: acquiring a speech training dataset, wherein the speech training dataset contains multiple training samples, each training sample containing: a speech sample signal, and label information corresponding to the speech sample signal, the label information being used to at least indicate the source distance and speech quality level corresponding to the speech sample signal; using an unsupervised pre-trained model to extract features from the speech sample signal to obtain the acoustic features corresponding to the speech sample signal; and training the speech quality assessment model and the source distance estimation model respectively based on the acoustic features and label information corresponding to the speech sample signal to obtain the speech signal processing model.

[0009] In some embodiments, obtaining a speech training dataset includes: obtaining a clean speech signal; using an acoustic simulation tool, simulating the generation of first speech signals corresponding to different sound source distances based on the clean speech signal; adding different types and signal-to-noise ratios of noise to the first speech signal to obtain speech sample signals with different speech quality levels; determining the label information corresponding to the speech sample signals, and constructing a speech training dataset based on the speech sample signals and the label information.

[0010] In some embodiments, a clean speech signal is a speech signal from an interference-free speech stream.

[0011] In some embodiments, the acoustic features include: a first acoustic feature and a second acoustic feature; feature extraction of the speech sample signal includes: using a first convolutional network layer in a pre-trained model to perform local time-frequency analysis on the waveform of the speech sample signal to obtain the first acoustic feature; the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, the first convolutional network layer being a convolutional network layer in the pre-trained model with a network depth less than a first preset depth; using a second convolutional network layer in the pre-trained model to perform feature modeling on the speech sample signal to obtain the second acoustic feature; the second acoustic feature includes at least one of the following: statistical regularity feature of harmonic spacing, dynamic change feature of formant structure, temporal correlation feature of broadband energy distribution, the second convolutional network layer being a convolutional network layer in the pre-trained model with a network depth greater than a second preset depth, the second preset depth being greater than the first preset depth.

[0012] In some embodiments, the acoustic features include: a first acoustic feature and a second acoustic feature; feature extraction of the original speech signal includes: using a first convolutional network layer in a pre-trained model to perform local time-frequency analysis on the waveform of the original speech signal to obtain the first acoustic feature; the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, the first convolutional network layer being a convolutional network layer in the pre-trained model with a network depth less than a first preset depth; using a second convolutional network layer in the pre-trained model to perform feature modeling on the original speech signal to obtain the second acoustic feature; the second acoustic feature includes at least one of the following: statistical regularity feature of harmonic spacing, dynamic change feature of formant structure, temporal correlation feature of broadband energy distribution, the second convolutional network layer being a convolutional network layer in the pre-trained model with a network depth greater than a second preset depth, the second preset depth being greater than the first preset depth.

[0013] In some embodiments, the speech quality assessment model includes: a convolutional module, a temporal modeling module, a temporal feature compression module, a linear layer, and an activation module; the convolutional module is used to further extract local spatial features of the acoustic features extracted by the pre-trained model; the temporal modeling module is used to capture the temporal sequence information between local spatial features through a bidirectional long short-term memory network and perform layer normalization on the features; the temporal feature compression module is used to perform global average pooling on the features output by the temporal modeling module; and the linear layer and activation module is used to map the features output by the temporal feature compression module using an activation function to obtain the first parameter corresponding to the acoustic feature. The method further includes: during model training, obtaining the first parameters determined by the speech quality assessment model based on the acoustic features corresponding to the speech sample signals; determining the first loss function value based on the first parameters and the first label information corresponding to the speech sample signals in the speech training dataset; the first loss function is used to characterize the error between the speech quality characterized by the first parameters and the speech quality characterized by the first label, the first label being determined based on the mean square error and phase-aware Mel loss between the Mel spectrum of the clean speech signal and the Mel spectrum of the speech sample signal corresponding to the clean speech signal; and adjusting the model parameters of the speech quality assessment model based on the first loss function value.

[0014] In some embodiments, the sound source distance estimation model includes: a global self-attention module, a global pooling layer, and multiple parallel linear output layers; the global self-attention module is used to determine the attention weights corresponding to the acoustic features extracted by the pre-trained model through an attention mechanism, and the attention weights are used to characterize the degree of influence of the acoustic features on judging the sound source distance; the global pooling layer is used to extract the target feature most relevant to the sound source distance from the acoustic features based on the attention weights; the linear output layers include: a first linear output layer, a second linear output layer, and a third linear output layer; the first linear output layer is used to determine a second parameter based on the target feature; the second linear output layer is used to determine the energy attenuation rate corresponding to the acoustic feature based on the target feature; and the third linear output layer is used to determine the high-frequency loss ratio corresponding to the acoustic feature based on the target feature. For example, the method also includes: during model training, obtaining the second parameter, energy attenuation rate, and high-frequency loss ratio determined by the sound source distance estimation model based on the acoustic features corresponding to the speech sample signals; determining the second loss function value based on the first error, the second error, and the third error; the first error is the error between the sound source distance represented by the second parameter and the sound source distance represented by the second label corresponding to the speech sample signals in the speech training dataset; the second error is the error between the energy attenuation rate and the energy attenuation label, the energy attenuation label being determined based on the speech energy of the clean speech signal and the speech energy of the speech sample signals corresponding to the clean speech signal; the third error is the error between the high-frequency loss ratio and the spectral loss label; and adjusting the model parameters of the sound source distance estimation model based on the second loss function value.

[0015] In some embodiments, determining the interfering speech stream in the original speech signal based on the first parameter and the second parameter includes: determining the first weight corresponding to the first parameter and the second weight corresponding to the second parameter; determining the speech validity score corresponding to the speech stream based on the first parameter, the first weight, the second parameter and the second weight; and determining the speech stream in the original speech signal whose speech validity score is less than a preset score threshold as the interfering speech stream.

[0016] In some embodiments, the speech validity score is used to characterize at least one of speech quality and the sound source distance.

[0017] According to another aspect of the embodiments of this application, a speech signal processing apparatus is also provided, comprising: a feature extraction module, configured to acquire an original speech signal and extract features from the original speech signal to obtain acoustic features corresponding to the original speech signal; a feature analysis module, configured to analyze the acoustic features using a speech signal processing model to obtain a first parameter and a second parameter corresponding to the original speech signal; the first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal, wherein the sound source distance is the distance between the sound source generating the speech stream and the sound pickup device acquiring the speech stream; and an interference filtering module, configured to determine the interference speech stream in the original speech signal based on the first parameter and the second parameter, and filter out the interference speech stream from the original speech signal to obtain a target speech signal.

[0018] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program executes a speech signal processing method during runtime.

[0019] According to another aspect of the embodiments of this application, a non-volatile storage medium is also provided, the non-volatile storage medium including a stored computer program, wherein the device where the non-volatile storage medium is located executes a speech signal processing method by running the computer program.

[0020] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of a speech signal processing method. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the disclosed drawings without creative effort.

[0022] Figure 1 is a hardware structure block diagram of a computer terminal (or electronic device) for implementing a method of speech signal processing according to an embodiment of this application;

[0023] Figure 2 is a schematic diagram of a speech signal processing method according to an embodiment of this application;

[0024] Figure 3 is a schematic diagram of the architecture of a joint estimation method for speech quality and speaker distance (sound source distance) based on an unsupervised pre-trained model provided according to an embodiment of this application;

[0025] Figure 4 is a schematic diagram of the process of a joint estimation method for speech quality and speaker distance (sound source distance) based on an unsupervised pre-trained model provided in an embodiment of this application;

[0026] Figure 5 is a schematic diagram of the network structure of a speech quality assessment model provided according to an embodiment of this application;

[0027] Figure 6 is a schematic diagram of the network structure of a sound source distance estimation model provided according to an embodiment of this application;

[0028] Figure 7 is a schematic diagram of a speech signal processing device according to an embodiment of this application. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] To facilitate a better understanding of the embodiments of this application by those skilled in the art, some technical terms or nouns involved in the embodiments of this application are explained as follows:

[0032] Unsupervised pre-trained model: A speech representation learning system based on a two-stage self-supervised framework of contrastive learning and quantization representation. It employs a contrastive loss function to construct context-dependent acoustic feature representations; through joint optimization via a multi-task learning mechanism, its hidden feature space simultaneously encompasses time-frequency features related to speech quality assessment and information such as energy and high-frequency attenuation required for speaker-speaker distance (sound source distance) estimation, providing a unified feature representation for downstream tasks.

[0033] Speech quality assessment: A quantitative method for speech intelligibility based on acoustic feature analysis. This method constructs a computational model to model key acoustic parameters such as fundamental frequency stability and formant sharpness, evaluating the semantic intelligibility of speech signals under specific acoustic conditions. The assessment system comprises two core dimensions: intelligibility, reflecting semantic transmission efficiency, and clarity, representing signal integrity.

[0034] Speaker distance (sound source distance) estimation: The spatial distance between the speaker (sound source) and the pickup device is estimated by combining information such as the required harmonic structure changes, energy attenuation, reverberation characteristics and high frequency attenuation.

[0035] In complex acoustic scenarios (such as conference rooms with multiple people talking, noisy public places, etc.), the VAD model of related technologies cannot effectively identify background human voices (other people's voices) and effective human voices (the user's target voice), resulting in a poor user voice interaction experience.

[0036] To address the aforementioned issues, this application provides relevant solutions, which are detailed below.

[0037] According to an embodiment of this application, a method embodiment for speech signal processing is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0038] The method embodiments provided in this application can be executed in a mobile terminal, computer terminal, or similar computing device. Figure 1 shows a hardware structure block diagram of a computer terminal (or electronic device) for implementing a voice signal processing method. As shown in Figure 1, the computer terminal 10 (or electronic device) may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), 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 a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that the structure shown in Figure 1 is only illustrative and does not limit the structure of the above-described electronic device. For example, the computer terminal 10 may also include more or fewer components than shown in Figure 1, or have a different configuration than shown in Figure 1.

[0039] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or electronic device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0040] 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 voice signal processing method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby implementing the aforementioned voice signal processing method. The memory 104 may include high-speed random access memory, and may also include 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 memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0042] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or electronic device).

[0043] Under the above operating environment, this application provides a speech signal processing method. Figure 2 is a schematic diagram of the speech signal processing method provided according to this application. As shown in Figure 2, the method includes the following steps:

[0044] Step S202: Obtain the original speech signal and extract features from the original speech signal to obtain the acoustic features corresponding to the original speech signal;

[0045] Step S204: Using a speech signal processing model, analyze the acoustic features to obtain the first parameter and the second parameter corresponding to the original speech signal. The first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal. The sound source distance is the distance between the sound source that generates the speech stream and the sound pickup device that collects the speech stream.

[0046] Step S206: Based on the first parameter and the second parameter, determine the interfering speech stream in the original speech signal, and filter out the interfering speech stream from the original speech signal to obtain the target speech signal.

[0047] By taking the above steps and jointly estimating the speech intelligibility of the speech signal and the distance to the sound source, the goal of filtering out background human voice interference in speech human-computer interaction systems in complex acoustic scenarios is achieved. This solves the technical problem that human-computer speech interaction is often affected by background noise in complex acoustic scenarios, resulting in poor user experience. It can effectively process speech signals in complex acoustic scenarios and improve the performance of speech human-computer interaction systems.

[0048] The speech signal processing method in steps S202 to S208 of the embodiments of this application will be further described below.

[0049] Figure 3 is a schematic diagram of the architecture of a joint estimation method for speech quality and speaker distance (sound source distance) based on an unsupervised pre-trained model according to an embodiment of this application. As shown in Figure 3, the speech signal processing model includes: a pre-trained model, a speech quality assessment model (for speech quality assessment), a sound source distance estimation model (for speaker distance estimation), and an adaptive fusion module. By quantitatively evaluating speech clarity and speaker distance (sound source distance), a comprehensive score (taking the range of 0 to 1 as an example) is output to filter background interference voices. The closer the comprehensive score is to 1, the clearer the speech and the closer the speaker is to the microphone. By setting a reasonable threshold, a cleaner input signal is provided to the downstream task speech recognition system, thereby improving the recognition accuracy and improving the user's voice interaction experience.

[0050] The construction and training workflow of the above architecture (speech signal processing model) is shown in Figure 4, and the details are as follows.

[0051] In some embodiments of this application, the training steps of the speech signal processing model include the following steps: obtaining a speech training dataset, wherein the speech training dataset contains multiple training samples, each training sample containing: a speech sample signal, and label information corresponding to the speech sample signal, the label information being used at least to indicate the source distance and speech quality level corresponding to the speech sample signal; using an unsupervised pre-trained model to extract features from the speech sample signal to obtain the acoustic features corresponding to the speech sample signal; and training the speech quality assessment model and the source distance estimation model respectively based on the acoustic features and label information corresponding to the speech sample signal to obtain the speech signal processing model.

[0052] First, construct the speech training dataset for model training, as follows.

[0053] In some embodiments of this application, obtaining a speech training dataset includes: obtaining a clean speech signal, wherein the clean speech signal is a speech signal of an interference-free speech stream; using an acoustic simulation tool, simulating the generation of first speech signals corresponding to different sound source distances based on the clean speech signal; adding noise of different types and signal-to-noise ratios to the first speech signal to obtain speech sample signals of different speech quality levels; determining the label information corresponding to the speech sample signals, and constructing a speech training dataset based on the speech sample signals and the label information.

[0054] Specifically, clean speech signals are obtained from public speech datasets, which typically contain high-quality speech recordings. Then, acoustic simulation tools (such as the Pyroomacoustics project) are used to simulate speech signals at different distances from the microphone based on the clean speech signals; these are the first speech signals corresponding to different sound source distances. Different types of noise with random signal-to-noise ratios are added to the first speech signals to construct speech signals of varying quality, resulting in speech sample signals. Simultaneously, the speech sample signals are labeled with corresponding tags. For example, a small number of speech sample signals can be labeled with MOS (Mean Opinion Score) subjective ratings, and the MOS scores can be normalized to the range of 0-1 for supervised fine-tuning training.

[0055] Once the speech training dataset is constructed, it can be used to train the speech signal processing models, namely the speech quality assessment model and the sound source distance estimation model.

[0056] It should be noted that in the model training and subsequent model application processes of this application embodiment, the speech sample signal / original speech signal is not directly analyzed. Instead, an unsupervised pre-trained model can be used to extract the acoustic features corresponding to the speech sample signal / original speech signal, as detailed below.

[0057] In some embodiments of this application, the acoustic features include: a first acoustic feature and a second acoustic feature; feature extraction of the speech sample signal / original speech signal includes the following steps: using a first convolutional network layer in a pre-trained model to perform local time-frequency analysis on the waveform of the speech sample signal / original speech signal to obtain the first acoustic feature, wherein the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, and the first convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth less than a first preset depth; using a second convolutional network layer in the pre-trained model to perform feature modeling on the speech sample signal / original speech signal to obtain the second acoustic feature, wherein the second acoustic feature includes at least one of the following: statistical regularity feature of harmonic spacing, dynamic change feature of formant structure, temporal correlation feature of broadband energy distribution, and the second convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth greater than a second preset depth, and the second preset depth is greater than the first preset depth.

[0058] In related technologies, the evaluation of voice commands typically relies on manually designed acoustic features, such as spectral centroid and zero-crossing rate. Their generalization ability is limited by empirical assumptions in feature engineering and the limited coverage of distorted scenarios. Similarly, the accuracy of single-channel speaker distance estimation is significantly degraded due to the complex coupling of environmental acoustic characteristics (such as reverberation time and noise type). However, in this embodiment, an unsupervised pre-trained model is used for self-supervised learning with large-scale unlabeled data, enabling the extraction of implicitly encoded generalized acoustic features, thereby effectively covering complex acoustic scenarios.

[0059] Specifically, in this embodiment, the shallow convolutional network (i.e., the first convolutional network layer mentioned above) of the pre-trained model (taking Wav2Vec 2.0 as an example) performs local time-frequency analysis on the original waveform through a learnable filter bank. Its function is similar to an adaptively optimized short-time Fourier transform (STFT), which can directly capture low-level acoustic features such as fundamental frequency (F0) and harmonic energy distribution (i.e., the first acoustic features mentioned above). The deep convolutional module (i.e., the second convolutional network layer mentioned above) can gradually model higher-order harmonic modes through a hierarchical abstraction mechanism, including: the statistical law of harmonic spacing, the dynamic change of formant structure, and the temporal correlation of broadband energy distribution, etc., which are second acoustic features. Based on pre-training with more than 20,000 hours of multi-scene speech data (including different signal-to-noise ratios and reverberation environments), the model can implicitly learn the statistical correlation between acoustic characteristics and physical propagation laws, providing a robust acoustic representation foundation for downstream tasks (such as speech quality assessment and speaker distance estimation).

[0060] After extracting the acoustic features corresponding to the speech sample signal / original speech signal through the pre-trained model, the extracted acoustic features can be input into the speech quality assessment model and the sound source distance estimation model for training or processing. The speech quality assessment model and the sound source distance estimation model are introduced below.

[0061] Figure 5 is a schematic diagram of the network structure of a speech quality assessment model according to an embodiment of this application. As shown in Figure 5, the speech quality assessment model includes: a convolutional module, a temporal modeling module, a temporal feature compression module, a linear layer, and an activation module (linear layer + sigmoid). The convolutional module is used to further extract local spatial features of the acoustic features extracted by the pre-trained model, and includes a layer of Convolutional Neural Network (CNN) and an Exponential Linear Unit (ELU) activation function. The temporal modeling module consists of a layer of Bidirectional Long Short-Term Memory (Bi-LSTM) network and layer normalization, used to capture the temporal sequence information between local spatial features through the Bi-LSTM network and perform layer normalization on the features. The temporal feature compression module uses global average pooling. Pooling (GAP) reduces dimensionality along the time axis to achieve efficient information compression, and is used for global average pooling of the features output by the temporal modeling module. The linear layer and activation module use activation functions to map the features output by the temporal feature compression module to obtain the first parameter corresponding to the acoustic features. For example, the linear layer and sigmoid activation function map the features to a range of 0 to 1; closer to 1 indicates clearer speech, and closer to 0 indicates poorer speech quality. The input to this model is the 1024-dimensional features of a pre-trained model (wav2vec 2.0), and the output is a speech quality score (i.e., the first parameter).

[0062] The training process of the speech quality assessment model is described below.

[0063] In some embodiments of this application, the method further includes the following steps: during model training, obtaining first parameters determined by the speech quality assessment model based on the acoustic features corresponding to the speech sample signals; determining a first loss function value based on the first parameters and the first label information corresponding to the speech sample signals in the speech training dataset, wherein the first loss function is used to characterize the error between the speech quality characterized by the first parameters and the speech quality characterized by the first label, and the first label is determined based on the mean square error and phase-aware Mel loss between the Mel spectrum of the clean speech signal and the Mel spectrum of the speech sample signal corresponding to the clean speech signal; and adjusting the model parameters of the speech quality assessment model based on the first loss function value.

[0064] Specifically, the speech quality assessment model uses the mean squared error (MSE) loss function for supervised training. Its training data labels are generated by constraining a negative correlation between speech quality scores and Mel spectrogram differences. Specifically, the Mel spectrograms of each pair of "clean" and "enhanced" speech (i.e., the Mel spectrogram of the clean speech signal and the Mel spectrogram of the corresponding speech sample signal) are extracted, and the Log-MSE (log mean squared error) and Phase-Aware Mel Loss (phase-aware Mel loss) between them are calculated. As shown in the following equation:

[0065] Among them, S clean and S enhanced These are the Mel spectra of clean speech (clean speech signal) and enhanced speech (speech sample signal corresponding to the clean speech signal), respectively. ∈ is a small constant used to avoid taking the logarithm of zero values, and N is the number of frames in the Mel spectrum. This represents the phase difference between clean speech and enhanced speech in the i-th frame.

[0066] Log-MSE focuses on the amplitude spectrum differences of speech signals, capturing the energy distribution. Phase-Aware Mel Loss, through phase information, can better capture the harmonic structure and temporal characteristics of speech. Combining the information from both, the two losses are weighted and combined to obtain a comprehensive difference score. Finally, the comprehensive difference score is normalized and used as the target label (i.e., the first label mentioned above).

[0067] The sound source distance estimation model is introduced below.

[0068] Figure 6 is a schematic diagram of the network structure of a sound source distance estimation model according to an embodiment of this application. As shown in Figure 6, the model input is the speech acoustic representation extracted by the pre-trained model (sharing features with the speech quality assessment model). The sound source distance estimation model includes: a global self-attention module, a global pooling layer, and multiple parallel linear output layers. The global self-attention module is used to determine the attention weights corresponding to the acoustic features extracted by the pre-trained model through an attention mechanism. The attention weights are used to characterize the degree of influence of the acoustic features on the judgment of the sound source distance. The global pooling layer is used to extract the target features most relevant to the sound source distance from the acoustic features based on the attention weights. The linear output layers include: a first linear output layer, a second linear output layer, and a third linear output layer. The first linear output layer is used to determine the second parameter based on the target feature. The second linear output layer is used to determine the energy attenuation rate corresponding to the acoustic feature based on the target feature. The third linear output layer is used to determine the high-frequency loss ratio corresponding to the acoustic feature based on the target feature.

[0069] The training process of the sound source distance estimation model is described below.

[0070] In some embodiments of this application, the method further includes the following steps: during model training, obtaining the second parameter, energy attenuation rate, and high-frequency loss ratio determined by the sound source distance estimation model based on the acoustic features corresponding to the speech sample signal; determining the second loss function value based on the first error, the second error, and the third error, wherein the first error is the error between the sound source distance represented by the second parameter and the sound source distance represented by the second label corresponding to the speech sample signal in the speech training dataset, the second error is the error between the energy attenuation rate and the energy attenuation label, the energy attenuation label is determined based on the speech energy of the clean speech signal and the speech energy of the speech sample signal corresponding to the clean speech signal, and the third error is the error between the high-frequency loss ratio and the spectrum loss label; and adjusting the model parameters of the sound source distance estimation model based on the second loss function value.

[0071] In this embodiment, the three parallel linear output layers in the sound source distance estimation model are used for different prediction tasks. Specifically, the main task uses a regression loss function L... distance The speaker distance score (i.e., the second parameter) is directly predicted, while two auxiliary tasks predict the energy attenuation rate and the high-frequency loss ratio, respectively. A multi-task loss function is designed by incorporating physical acoustic laws (taking 16kHz audio as an example, the energy attenuation during sound wave propagation follows the inverse square law, and the high-frequency components (4kHz to 8kHz) show significant attenuation characteristics with increasing distance) to jointly constrain the physical consistency of the model.

[0072] Specifically, the energy decay regression loss L energy (The second error) is based on time-domain energy calculation to enhance the speech segment energy E.enhanced With pure voice energy E clean The ratio is used as a label to constrain the predicted decay rate output by the model. High-frequency loss proportional regression loss L high-frequency (The third error) is based on the ratio of the sum of energy within a specific frequency band to the total spectral energy, which serves as a label to constrain the predicted high-frequency loss ratio output by the model. This is illustrated in the following formula.

[0073] Where M represents the number of frames in the speech segment, N represents the number of sampling points per frame, and x m (n) represents the value of the nth sample point in the mth frame.

[0074] After obtaining the first parameter output by the speech quality assessment model and the second parameter output by the sound source distance estimation model, the first and second parameters can be dynamically weighted and fused using the adaptive fusion module. The specific steps are as follows.

[0075] In some embodiments of this application, determining the interfering speech stream in the original speech signal based on the first parameter and the second parameter includes the following steps: determining the first weight corresponding to the first parameter and the second weight corresponding to the second parameter; determining the speech validity score corresponding to the speech stream based on the first parameter, the first weight, the second parameter and the second weight, wherein the speech validity score is used to characterize at least one of speech quality and the sound source distance; and determining the speech stream in the original speech signal with a speech validity score less than a preset score threshold as the interfering speech stream.

[0076] Specifically, the adaptive fusion module receives the speech quality score (first parameter) from the speech quality assessment model and the speaker distance information (second parameter) from the sound source distance estimation model. It then uses a learnable fully connected layer to dynamically weight and fuse cross-dimensional features, generating a final comprehensive speech evaluation score (speech validity score). This score can be used to filter out interfering speech. The module employs an end-to-end training strategy, using speech data with subjective MOS annotations as supervision signals. By minimizing the mean square error (MSE) between the predicted score and the manually annotated MOS values, it ensures consistency between speech quality perception and physical acoustic laws in multi-dimensional evaluation.

[0077] This application uses an unsupervised pre-trained model to extract features from the base model: through large-scale data learning, it can extract acoustic features rich in information, including time-frequency characteristics, harmonic structure, energy attenuation, reverberation characteristics, etc.; it jointly optimizes speech quality assessment and speaker distance (sound source distance) estimation, designs a multi-task loss function by incorporating physical acoustic laws, and jointly constrains the physical consistency of the model; it designs a learnable weight allocation network, dynamically fuses the outputs of the speech quality and speaker distance sub-networks, and generates a final comprehensive score. Through the above methods, in noisy environments, it can help VAD further filter human voices of varying degrees, providing a cleaner speech signal for the speech recognition system, reducing the confusion caused by misidentification due to background human voices, and significantly improving the user's voice interaction experience.

[0078] According to an embodiment of this application, an embodiment of a speech signal processing device is also provided. Figure 7 is a schematic diagram of the structure of a speech signal processing device according to an embodiment of this application. As shown in Figure 7, the device includes:

[0079] The feature extraction module 70 is used to acquire the original speech signal and extract features from the original speech signal to obtain the acoustic features corresponding to the original speech signal.

[0080] The feature analysis module 72 is used to analyze acoustic features using a speech signal processing model to obtain the first parameter and the second parameter corresponding to the original speech signal. The first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal. The sound source distance is the distance between the sound source that generates the speech stream and the sound pickup device that collects the speech stream.

[0081] The interference filtering module 74 is used to determine the interference speech stream in the original speech signal based on the first parameter and the second parameter, and filter out the interference speech stream from the original speech signal to obtain the target speech signal.

[0082] In some embodiments, the speech signal processing model includes: a pre-trained model, a speech quality assessment model, and a source distance estimation model; the device further includes a model training module for: acquiring a speech training dataset, wherein the speech training dataset contains multiple training samples, each training sample containing: a speech sample signal, and label information corresponding to the speech sample signal, the label information being used to at least indicate the source distance and speech quality level corresponding to the speech sample signal; using an unsupervised pre-trained model to extract features from the speech sample signal to obtain the acoustic features corresponding to the speech sample signal; and training the speech quality assessment model and the source distance estimation model respectively based on the acoustic features and label information corresponding to the speech sample signal to obtain the speech signal processing model.

[0083] In some embodiments, the model training module is further configured to: acquire clean speech signals, wherein the clean speech signals are speech signals of interference-free speech streams; use acoustic simulation tools to simulate and generate first speech signals corresponding to different sound source distances based on the clean speech signals; add noise of different types and signal-to-noise ratios to the first speech signals to obtain speech sample signals of different speech quality levels; determine the label information corresponding to the speech sample signals, and construct a speech training dataset based on the speech sample signals and the label information.

[0084] In some embodiments, the acoustic features include: a first acoustic feature and a second acoustic feature; the model training module is further configured to: use a first convolutional network layer in the pre-trained model to perform local time-frequency analysis on the waveform of the speech sample signal to obtain the first acoustic feature, wherein the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, and the first convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth less than a first preset depth; use a second convolutional network layer in the pre-trained model to perform feature modeling on the speech sample signal to obtain the second acoustic feature, wherein the second acoustic feature includes at least one of the following: statistical regularity feature of harmonic spacing, dynamic change feature of formant structure, temporal correlation feature of broadband energy distribution, and the second convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth greater than a second preset depth, and the second preset depth is greater than the first preset depth.

[0085] In some embodiments, the acoustic features include: a first acoustic feature and a second acoustic feature; the feature extraction module 70 is further configured to: use a first convolutional network layer in the pre-trained model to perform local time-frequency analysis on the waveform of the original speech signal to obtain the first acoustic feature, wherein the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, and the first convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth less than a first preset depth; use a second convolutional network layer in the pre-trained model to perform feature modeling on the original speech signal to obtain the second acoustic feature, wherein the second acoustic feature includes at least one of the following: statistical regularity feature of harmonic spacing, dynamic change feature of formant structure, temporal correlation feature of broadband energy distribution, and the second convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth greater than a second preset depth, and the second preset depth is greater than the first preset depth.

[0086] In some embodiments, the speech quality assessment model includes: a convolutional module, a temporal modeling module, a temporal feature compression module, a linear layer, and an activation module. The convolutional module further extracts local spatial features of the acoustic features extracted by the pre-trained model; the temporal modeling module captures the temporal sequence information between local spatial features using a bidirectional long short-term memory network and performs layer normalization on the features; the temporal feature compression module performs global average pooling on the features output by the temporal modeling module; and the linear layer and activation module uses an activation function to map the features output by the temporal feature compression module to obtain the first parameter corresponding to the acoustic features. The processing device is further configured to: during model training, acquire first parameters determined by the speech quality assessment model based on the acoustic features corresponding to the speech sample signals; determine a first loss function value based on the first parameters and the first label information corresponding to the speech sample signals in the speech training dataset, wherein the first loss function is used to characterize the error between the speech quality characterized by the first parameters and the speech quality characterized by the first label, and the first label is determined based on the mean square error and phase-aware Mel loss between the Mel spectrum of the clean speech signal and the Mel spectrum of the speech sample signal corresponding to the clean speech signal; and adjust the model parameters of the speech quality assessment model based on the first loss function value.

[0087] In some embodiments, the sound source distance estimation model includes: a global self-attention module, a global pooling layer, and multiple parallel linear output layers. The global self-attention module determines the attention weights corresponding to the acoustic features extracted by the pre-trained model through an attention mechanism. These attention weights characterize the influence of the acoustic features on the sound source distance assessment. The global pooling layer extracts the target feature most relevant to the sound source distance from the acoustic features based on the attention weights. The linear output layers include: a first linear output layer, a second linear output layer, and a third linear output layer. The first linear output layer determines a second parameter based on the target feature. The second linear output layer determines the energy attenuation rate corresponding to the acoustic feature based on the target feature, and the third linear output layer determines the high-frequency loss ratio corresponding to the acoustic feature based on the target feature. The audio signal processing device is also used to: during model training, acquire the second parameter, energy attenuation rate, and high-frequency loss ratio determined by the sound source distance estimation model based on the acoustic features corresponding to the speech sample signals; determine the second loss function value based on the first error, the second error, and the third error, wherein the first error is the error between the sound source distance represented by the second parameter and the sound source distance represented by the second label corresponding to the speech sample signals in the speech training dataset, the second error is the error between the energy attenuation rate and the energy attenuation label, the energy attenuation label is determined based on the speech energy of the clean speech signal and the speech energy of the speech sample signals corresponding to the clean speech signal, and the third error is the error between the high-frequency loss ratio and the spectral loss label; and adjust the model parameters of the sound source distance estimation model based on the second loss function value.

[0088] In some embodiments, the interference filtering module 74 is further configured to: determine a first weight corresponding to a first parameter and a second weight corresponding to a second parameter; determine a speech validity score corresponding to a speech stream based on the first parameter, the first weight, the second parameter, and the second weight, wherein the speech validity score is used to characterize at least one of speech quality and the sound source distance; and determine speech streams in the original speech signal whose speech validity score is less than a preset score threshold as interference speech streams.

[0089] It should be noted that each module in the above-mentioned speech signal processing device can be a program module (for example, a set of program instructions to implement a certain function) or a hardware module. For the latter, it can be manifested in the following forms, but is not limited to them: each of the above modules is manifested as a processor, or the functions of each of the above modules are implemented by a processor.

[0090] It should be noted that the speech signal processing device provided in this embodiment can be used to execute the speech signal processing method shown in FIG2. Therefore, the relevant explanations and descriptions of the above speech signal processing method also apply to the embodiments of this application, and will not be repeated here.

[0091] This application embodiment also provides a non-volatile storage medium, which includes a stored computer program. The device containing the non-volatile storage medium executes the following speech signal processing method by running the computer program: acquiring an original speech signal and extracting features from the original speech signal to obtain acoustic features corresponding to the original speech signal; using a speech signal processing model to analyze the acoustic features to obtain a first parameter and a second parameter corresponding to the original speech signal, wherein the first parameter characterizes the speech quality of the speech stream in the original speech signal, and the second parameter characterizes the distance to the sound source corresponding to the speech stream in the original speech signal, the sound source distance being the distance between the sound source generating the speech stream and the sound pickup device acquiring the speech stream; based on the first parameter and the second parameter, determining the interfering speech stream in the original speech signal and filtering out the interfering speech stream from the original speech signal to obtain the target speech signal.

[0092] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the speech signal processing method described in various embodiments of this application: acquiring an original speech signal and extracting features from the original speech signal to obtain acoustic features corresponding to the original speech signal; using a speech signal processing model to analyze the acoustic features to obtain a first parameter and a second parameter corresponding to the original speech signal, wherein the first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal, the sound source distance being the distance between the sound source generating the speech stream and the sound pickup device acquiring the speech stream; based on the first parameter and the second parameter, determining the interfering speech stream in the original speech signal and filtering out the interfering speech stream from the original speech signal to obtain the target speech signal.

[0093] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0094] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0096] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0097] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0098] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A speech signal processing method, comprising: The original speech signal is acquired, and features are extracted from the original speech signal to obtain the acoustic features corresponding to the original speech signal. A speech signal processing model is used to analyze the acoustic features to obtain a first parameter and a second parameter corresponding to the original speech signal. The first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal. The sound source distance is the distance between the sound source that generates the speech stream and the sound pickup device that acquires the speech stream. Based on the first parameter and the second parameter, the interfering speech stream in the original speech signal is determined, and the interfering speech stream is filtered out from the original speech signal to obtain the target speech signal.

2. The method according to claim 1, wherein, The speech signal processing model includes: a pre-trained model, a speech quality assessment model, and a sound source distance estimation model; the training steps of the speech signal processing model include: Obtain a speech training dataset; the speech training dataset contains multiple training samples, each of the training samples contains: a speech sample signal, and label information corresponding to the speech sample signal, the label information being used at least to indicate the sound source distance and speech quality level corresponding to the speech sample signal; The unsupervised pre-trained model is used to extract features from the speech sample signal to obtain the acoustic features corresponding to the speech sample signal. Based on the acoustic features and label information corresponding to the speech sample signal, the speech quality assessment model and the sound source distance estimation model are trained respectively to obtain the speech signal processing model.

3. The method according to claim 2, wherein, Obtaining the speech training dataset includes: Obtain a clean speech signal; Using acoustic simulation tools, based on the clean speech signal, first speech signals corresponding to different distances from the sound source are simulated and generated. Adding different types and signal-to-noise ratios of noise to the first speech signal yields speech sample signals with different speech quality levels; The label information corresponding to the speech sample signal is determined, and the speech training dataset is constructed based on the speech sample signal and the label information.

4. The method according to claim 3, wherein, The clean speech signal is a speech signal without interference from the speech stream.

5. The method according to claim 2, wherein, The acoustic features include: a first acoustic feature and a second acoustic feature; feature extraction of the speech sample signal includes: Using the first convolutional network layer in the pre-trained model, local time-frequency analysis is performed on the waveform of the speech sample signal to obtain a first acoustic feature; the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, and the first convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth less than a first preset depth; The second convolutional network layer in the pre-trained model is used to perform feature modeling on the speech sample signal to obtain the second acoustic feature; the second acoustic feature includes at least one of the following: statistical regularity features of harmonic spacing, dynamic change features of formant structure, and temporal correlation features of broadband energy distribution. The second convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth greater than a second preset depth, and the second preset depth is greater than the first preset depth.

6. The method according to claim 2, wherein, The acoustic features include: a first acoustic feature and a second acoustic feature; feature extraction of the original speech signal includes: Using the first convolutional network layer in the pre-trained model, local time-frequency analysis is performed on the waveform of the original speech signal to obtain a first acoustic feature; the first acoustic feature includes at least one of the following: fundamental frequency feature, harmonic energy distribution feature, and the first convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth less than a first preset depth; The second convolutional network layer in the pre-trained model is used to perform feature modeling on the original speech signal to obtain the second acoustic feature; the second acoustic feature includes at least one of the following: statistical regularity features of harmonic spacing, dynamic change features of formant structure, and temporal correlation features of broadband energy distribution. The second convolutional network layer is a convolutional network layer in the pre-trained model with a network layer depth greater than a second preset depth, and the second preset depth is greater than the first preset depth.

7. The method according to claim 2, wherein, The speech quality assessment model includes: a convolutional module, a temporal modeling module, a temporal feature compression module, a linear layer, and an activation module. The convolutional module further extracts local spatial features of the acoustic features extracted by the pre-trained model. The temporal modeling module captures the temporal sequence information between the local spatial features using a bidirectional long short-term memory network and performs layer normalization on the features. The temporal feature compression module performs global average pooling on the features output by the temporal modeling module. The linear layer and activation module uses an activation function to map the features output by the temporal feature compression module to obtain the first parameter corresponding to the acoustic features. The method further includes: During model training, the first parameter determined by the speech quality assessment model based on the acoustic features corresponding to the speech sample signal is obtained; Based on the first parameter and the first label information corresponding to the speech sample signal in the speech training dataset, a first loss function value is determined; the first loss function is used to characterize the error between the speech quality characterized by the first parameter and the speech quality characterized by the first label information, and the first label information is determined based on the mean square error and phase-aware Mel loss between the Mel spectrum of the clean speech signal and the Mel spectrum of the speech sample signal corresponding to the clean speech signal. Based on the first loss function value, the model parameters of the speech quality assessment model are adjusted.

8. The method according to claim 2, wherein, The sound source distance estimation model includes: a global self-attention module, a global pooling layer, and multiple parallel linear output layers; the global self-attention module is used to determine the attention weights corresponding to the acoustic features extracted by the pre-trained model through an attention mechanism, and the attention weights are used to characterize the degree of influence of the acoustic features on judging the sound source distance; the global pooling layer is used to extract the target feature most relevant to the sound source distance from the acoustic features based on the attention weights; the linear output layers include: a first linear output layer, a second linear output layer, and a third linear output layer; the first linear output layer is used to determine the second parameter based on the target feature; the second linear output layer is used to determine the energy attenuation rate corresponding to the acoustic feature based on the target feature; the second linear output layer is used to determine the high-frequency loss ratio corresponding to the acoustic feature based on the target feature; the method further includes: During model training, the second parameter, the energy attenuation rate, and the high-frequency loss ratio determined by the sound source distance estimation model based on the acoustic features corresponding to the speech sample signal are obtained. The second loss function value is determined based on the first error, the second error, and the third error; the first error is the error between the sound source distance represented by the second parameter and the sound source distance represented by the second label corresponding to the speech sample signal in the speech training dataset; the second error is the error between the energy attenuation rate and the energy attenuation label, the energy attenuation label being determined based on the speech energy of the clean speech signal and the speech energy of the speech sample signal corresponding to the clean speech signal; and the third error is the error between the high-frequency loss ratio and the spectrum loss label. Based on the second loss function value, the model parameters of the sound source distance estimation model are adjusted.

9. The method according to claim 1, wherein, Based on the first parameter and the second parameter, the interference speech stream in the original speech signal is determined as follows: Determine the first weight corresponding to the first parameter and the second weight corresponding to the second parameter; Based on the first parameter, the first weight, the second parameter, and the second weight, the speech validity score corresponding to the speech stream is determined. The speech streams in the original speech signal whose speech validity score is less than a preset score threshold are identified as the interference speech streams.

10. The method according to claim 9, wherein, The speech validity score is used to characterize at least one of speech quality and the sound source distance.

11. A speech signal processing apparatus, comprising: The feature extraction module is used to acquire the original speech signal and extract features from the original speech signal to obtain the acoustic features corresponding to the original speech signal. The feature analysis module is used to analyze the acoustic features using a speech signal processing model to obtain a first parameter and a second parameter corresponding to the original speech signal. The first parameter is used to characterize the speech quality of the speech stream in the original speech signal, and the second parameter is used to characterize the distance of the sound source corresponding to the speech stream in the original speech signal. The sound source distance is the distance between the sound source that generates the speech stream and the sound pickup device that acquires the speech stream. An interference filtering module is used to determine the interference speech stream in the original speech signal based on the first parameter and the second parameter, and to filter out the interference speech stream from the original speech signal to obtain the target speech signal.

12. The apparatus according to claim 11, wherein, The speech signal processing model includes: a pre-trained model, a speech quality assessment model, and a sound source distance estimation model; the device also includes a model training module, used for: Obtain a speech training dataset; the speech training dataset contains multiple training samples, each of the training samples contains: a speech sample signal, and label information corresponding to the speech sample signal, the label information being used at least to indicate the sound source distance and speech quality level corresponding to the speech sample signal; The unsupervised pre-trained model is used to extract features from the speech sample signal to obtain the acoustic features corresponding to the speech sample signal. Based on the acoustic features and label information corresponding to the speech sample signal, the speech quality assessment model and the sound source distance estimation model are trained respectively to obtain the speech signal processing model.

13. An electronic device, comprising: A memory and a processor, the processor being used to run a program stored in the memory; When the program is run, it performs the steps of the speech signal processing method according to any one of claims 1 to 10.

14. A non-volatile storage medium comprising a stored computer program, wherein the device containing the non-volatile storage medium executes the steps of the speech signal processing method according to any one of claims 1 to 10 by running the computer program.

15. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the speech signal processing method according to any one of claims 1 to 10.