Speech signal noise reduction method, system and device

Abstract

The present invention relates to the field of speech signal processing technology, and in particular to a speech signal denoising method, system and device. The method comprises the following steps: adaptively framing speech signal data to be processed to obtain framed speech signal data; semantically embedding space processing is performed according to the framed speech signal data to obtain semantically embedded space data; semantic vector field is constructed according to the semantically embedded space data to generate speech semantic vector field data; speech noise area is marked on the speech semantic vector field data to generate semantic noise area data; noise frame repair processing is performed according to the semantic noise area data to generate denoised speech signal data; auditory masking spectrum correction is performed on the denoised speech signal data to generate enhanced speech signal data. The present invention accurately guides speech signal denoising through semantic information, and can still achieve excellent denoising effect even in complex scenarios such as low signal-to-noise ratio, significantly improving user experience.

Description

Technical Field

[0001] The present invention relates to the technical field of speech signal processing, and in particular to a speech signal noise reduction method, system and device. Background Art

[0002] With the rapid development of information technology, voice communication and speech recognition technologies have been widely used in various fields. In particular, in scenarios such as smartphones, smart assistants, and remote conferences, clear voice signal transmission and processing are crucial. However, in practical applications, voice signals are often interfered with by various noises, which can come from the environment (such as traffic noise, crowd noise, etc.) or the equipment itself (such as poor microphone quality, unstable signal transmission, etc.). These noises can significantly reduce the clarity and intelligibility of voice signals, and even cause information transmission errors. However, traditional voice signal denoising methods mainly include spectral subtraction, Wiener filtering, Kalman filtering, etc. These methods can reduce the interference of background noise to a certain extent, but often have problems such as unsatisfactory noise reduction effect and voice distortion. Especially when processing low signal-to-noise ratio audio, it is easy to over-suppress weak voice signals, making it difficult to hear the voice clearly, affecting the user experience. Summary of the invention

[0003] Based on this, the present invention provides a speech signal noise reduction method, system and device to solve at least one of the above technical problems.

[0004] To achieve the above object, a method for reducing noise of a speech signal comprises the following steps:

[0005] Step S1: Adaptively frame the speech signal data to be processed to obtain framed speech signal data; perform fundamental frequency voiceprint feature fusion according to the framed speech signal data to generate initial voiceprint base data;

[0006] Step S2: Perform acoustic context enhancement processing on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform temporal semantic trajectory processing on the semantic embedding space data to obtain speech semantic trajectory data;

[0007] Step S3: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; marking the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data;

[0008] Step S4: performing noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; performing key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; performing noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data;

[0009] Step S5: performing multi-base frequency spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; performing auditory masking spectrum correction based on the multi-base frequency speech spectrum data, and performing speech signal reconstruction processing to generate enhanced speech signal data.

[0010] The present invention performs adaptive framing processing on the speech signal to be processed, and can flexibly adjust the analysis window according to the dynamic characteristics of the speech signal, thereby ensuring the accurate capture of different frequency components. This framing processing can capture the instantaneous characteristics of the speech signal, which is helpful to extract the basic frequency characteristics of the speech signal. The fusion of the fundamental frequency voiceprint features not only extracts the unique information closely related to the identity of the speaker in the speech signal, but also enhances the robustness to environmental changes. Using the initial voiceprint base data for acoustic context enhancement processing helps to improve the quality of the acoustic feature vector data. This enhancement processing can effectively consider the context information and reduce the interference of external noise on the speech signal. By performing semantic embedding space processing on the framed speech signal data through the acoustic feature vector data, the speech signal can be mapped to a high-dimensional semantic space, thereby making the semantic information of the speech signal clearer. Based on the speech semantic trajectory data and the semantic embedding space data, the semantic vector field is constructed to generate a combination of acoustic information and semantic information. By accurately marking the noise area, the direction and strategy of noise suppression can be effectively guided, and the accuracy of noise reduction processing can be improved. By using the semantic noise region data to perform noise suppression on the speech semantic vector field data, the noise component in the speech signal can be effectively reduced, thereby improving the clarity of the speech signal. By using the noise reduction semantic vector field data to perform key semantic noise frame recognition on the speech semantic trajectory data, the noise frames that have a greater impact on speech understanding can be effectively identified. By performing noise frame repair processing on the speech key noise frame data, the identified noise frames can be repaired, which can significantly improve the audibility and comprehensibility of the speech signal. By performing multi-base spectrum decomposition processing on the noise reduction speech signal data, the speech signal can be decomposed into multiple frequency components, making its characteristics in the frequency domain more obvious. By performing auditory masking spectrum correction on the multi-base frequency speech spectrum data, the spectral quality of the speech signal can be further improved, and unnecessary frequency interference can be eliminated. On this basis, the speech signal reconstruction processing can be performed to generate enhanced speech signal data. The processed frequency domain information is effectively converted into a high-quality time domain speech signal, thereby improving the overall quality of the speech signal and the user experience. Therefore, a speech signal denoising method of the present invention maps the semantic information of the speech signal to a high-dimensional space to form semantic trajectory data, performs semantic-guided noise recognition on the speech signal according to the semantic trajectory data, obtains key noise speech frames, considers the acoustic characteristics of the key noise speech frames, and performs precise noise reduction and restoration in combination with speech trends, thereby ensuring that the semantic information of the speech is retained during the enhancement process, improving the clarity and intelligibility of the speech, and effectively avoiding the problem of damage to the speech caused by excessive suppression.

[0011] Preferably, step S1 comprises the following steps:

[0012] Step S11: performing voice endpoint detection on the voice signal data to be processed and performing silent segment elimination processing to generate initial voice signal data;

[0013] Step S12: extracting the fundamental frequency period according to the initial speech signal data to obtain fundamental frequency contour feature data;

[0014] Step S13: performing adaptive framing processing according to the initial voice signal data to obtain framed voice signal data;

[0015] Step S14: performing fast Fourier transform processing on the framed speech signal data, and calculating the Mel frequency cepstrum coefficients according to the Mel filter to generate Mel frequency spectrum feature matrix data;

[0016] Step S15: performing dynamic time warping processing on the fundamental frequency profile feature data through the Mel frequency spectrum feature matrix data, and performing fundamental frequency voiceprint feature fusion to obtain fundamental frequency fusion feature data;

[0017] Step S16: Estimating the voiceprint distribution probability based on the baseband fusion feature data, and performing Gaussian fuzzy generalization processing to generate initial voiceprint base data.

[0018] The present invention performs voice endpoint detection on the voice signal data to be processed, and can accurately identify the start and end positions of the voice, thereby effectively eliminating the silent segment. The fundamental frequency period extraction is performed according to the initial voice signal data, and the basic frequency information of the voice signal can be extracted. The initial voice signal data is adaptively framed, and the continuous voice signal can be divided into several short time frames, which can not only capture the detailed changes of the voice, but also adapt to the changes in the voice rate in different scenarios, thereby improving the flexibility and effectiveness of the processing. The framed voice signal data is processed by fast Fourier transform, and the time domain signal can be converted into a frequency domain signal, thereby revealing the spectral characteristics of the voice signal. Combined with the Mel filter to calculate the Mel frequency cepstrum coefficient (MFCC), the Mel spectrum feature matrix can more accurately reflect the perception characteristics of the human ear to different frequencies, and improve the representativeness and effectiveness of the voice features. The fundamental frequency contour feature data is dynamically time-warped by the Mel spectrum feature matrix data, which can effectively align the time differences between different voice samples, and ensure the consistency of the features even when there are differences in the speaking speed. Through fuzzy generalization processing, the noise influence of the feature data can be effectively reduced, and the robustness and generalization ability of the voiceprint can be improved.

[0019] Preferably, step S2 comprises the following steps:

[0020] Step S21: performing acoustic context enhancement processing according to the initial voiceprint base data to generate acoustic feature vector data;

[0021] Step S22: performing frame-by-frame speech recognition on the framed speech signal data using a preset speech recognition model based on the acoustic feature vector data to obtain frame-level speech text data;

[0022] Step S23: performing semantic embedding space processing according to the frame-level speech text data to obtain semantic embedding space data;

[0023] Step S24: semantic space mapping is performed on the framed speech signal data by using the semantic embedding space data to generate framed speech semantic coordinate data;

[0024] Step S25: performing time-series semantic trajectory processing according to the framed speech semantic coordinate data to obtain speech semantic trajectory data.

[0025] The present invention performs acoustic context enhancement processing based on the initial voiceprint base data, which can effectively improve the quality and accuracy of the acoustic feature vector data, improve the robustness of the features extracted from the initial voiceprint base data to environmental changes, and enhance the expressiveness of the speech information in the speech signal. Using a preset speech recognition model to perform frame-by-frame speech recognition on the framed speech signal data can achieve accurate speech-to-text processing, improve the accuracy of speech recognition, ensure that the content of each frame of the speech signal is effectively captured, and reduce the possibility of recognition errors caused by environmental noise. According to the frame-level speech text data, semantic embedding space processing is performed, and the text information can be mapped to a high-dimensional semantic space, which can better reflect the semantic characteristics of the speech content and improve the accuracy and depth of the understanding of the speech signal. By performing semantic space mapping on the framed speech signal data through semantic embedding space data, the semantic information of each frame of the speech signal can be represented in coordinate form, making the semantic relationship between different frames more intuitive. The temporal semantic trajectory processing based on the framed speech semantic coordinate data can reflect the semantic dynamics of speech signals over time and reveal the semantic change pattern of speech content over time. It not only reflects the meaning of individual words, but also captures the logical structure and semantic flow of the entire speech sequence.

[0026] Preferably, step S23 includes the following steps:

[0027] Step S231: converting the frame-level speech text data into a high-dimensional semantic vector to generate semantic original vector data;

[0028] Step S232: constructing a high-dimensional manifold structure according to the semantic original vector data to obtain semantic manifold structure data;

[0029] Step S233: using the frame-level speech text data to identify the core semantic factors of the semantic manifold structure data, and generating core semantic factor data;

[0030] Step S234: using a preset speech corpus to perform contextual semantic embedding matching on the frame-level speech text data, and generating semantic corpus through a generative adversarial network to obtain diversified corpus data;

[0031] Step S235: decoupling semantic features according to the diversified corpus data, and performing semantic association network processing to generate semantic relationship network data;

[0032] Step S236: Dynamically semantically fuse the semantic manifold structure data through the semantic relationship network data to obtain dynamic semantic embedding data;

[0033] Step S237: semantically calibrate the dynamic semantic embedding data, and construct a KD tree fast index based on the core semantic factor data, so as to obtain semantic embedding space data.

[0034] The present invention converts text information into a high-dimensional vector form, making the expression of semantic information richer and more comprehensive. High-dimensional semantic vectors can capture subtle differences and potential semantic relationships in text. By constructing a high-dimensional manifold, the relationship and distribution characteristics between semantic vectors can be effectively represented. By identifying the core semantic factors of semantic manifold structure data with frame-level speech text data, key semantic factors can be extracted from complex semantic manifolds, helping to clarify the main information and themes in the speech signal, thereby enhancing the efficiency and accuracy of speech recognition. Through context matching, a diversified corpus related to speech content can be generated, enriching the background information of speech processing. By decoupling semantic features, the independence and mutual relationship between different semantic components can be better understood. Dynamic semantic fusion of semantic manifold structure data through semantic relationship network data realizes dynamic adjustment and optimization of semantic representation, ensuring that semantic embedding not only reflects the immediate semantics of the current speech signal, but also integrates historical context and potential semantic development trends, thereby improving the accuracy and adaptability of semantic expression. Through semantic scale calibration and KD tree fast index construction, not only the dynamic semantic embedding data is standardized, so that different semantic features can be compared and retrieved on a unified scale, but also efficient data indexing and query are achieved using KD tree.

[0035] Preferably, step S3 comprises the following steps:

[0036] Step S31: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data;

[0037] Step S32: performing grid tensor feature analysis on the speech semantic vector field data to generate grid semantic tensor data; performing semantic vector streamline tracing on the speech semantic vector field data through the grid semantic tensor data to obtain semantic vector streamline data;

[0038] Step S33: performing streamline stability evaluation according to the semantic vector streamline data to generate streamline stability evaluation data;

[0039] Step S34: performing semantic eddy current detection on the speech semantic vector field data based on the semantic vector streamline data to generate speech eddy current area data;

[0040] Step S35: calculating the average speed of the semantic vector streamline data through a preset multi-scale sliding window to generate semantic streamline speed sequence data;

[0041] Step S36: extracting semantic acceleration from speech eddy current region data based on semantic streamline velocity sequence data, and quantifying eddy current intensity to generate eddy current intensity data;

[0042] Step S37: Mark the speech noise area of ​​the speech semantic vector field data using the eddy current intensity data, the speech eddy current area data, and the streamline stability evaluation data to generate semantic noise area data.

[0043] The present invention constructs a semantic vector field according to speech semantic trajectory data and semantic embedding space data, represents the semantic information of speech signals in the form of space vectors, and provides a comprehensive understanding of speech content. By performing grid tensor feature analysis on speech semantic vector field data, important features in the semantic vector field can be extracted, and the structural features of speech signals can be revealed. By tracking semantic vector streamlines on speech semantic vector field data through grid semantic tensor data, the flow and change of semantic information in space can be displayed, which helps to identify the dynamic characteristics of speech signals. By analyzing the semantic vector streamline data, the continuity and consistency of the streamlines can be judged, which helps to identify the interruption or irregular change of the streamlines caused by noise interference. By eddy current detection, complex patterns and potential noise areas in speech signals can be identified. Speech eddy current area data helps to determine the part of the speech signal where interference exists. The calculation of streamline velocity can reveal the changing characteristics of speech signals on different time scales. By extracting semantic acceleration, the changing trend of speech signals can be analyzed, and quantifying eddy current intensity helps to evaluate the impact of noise on speech signals. By comprehensively considering eddy current intensity and streamline stability, the noise area in the speech signal can be accurately identified and marked.

[0044] Preferably, step S31 includes the following steps:

[0045] Step S311: performing semantic space gridding processing on the semantic embedding space data to generate grid semantic space data; performing initial semantic potential energy value calculation on the grid semantic space data to generate initial semantic potential energy data;

[0046] Step S312: constructing an initial vector field for the grid semantic space data using the initial semantic potential data to obtain initial semantic vector field data;

[0047] Step S313: dynamically adjusting the initial semantic vector field data using the speech semantic trajectory data to generate dynamic semantic vector field data; wherein step S313 is specifically as follows:

[0048] Step S3131: using the speech semantic trajectory data to perform semantic trajectory mapping processing on the grid semantic space data to generate semantic trajectory grid sequence data;

[0049] Step S3132: using the grid semantic space data to perform grid elimination on the semantic trajectory grid sequence data to obtain non-semantic trajectory grid data;

[0050] Step S3133: adjusting the semantic trajectory vector of the initial semantic vector field data using the speech semantic trajectory data based on the semantic trajectory grid sequence data to generate semantic trajectory vector adjustment data;

[0051] Step S3134: Based on the non-semantic trajectory grid data, the semantic trajectory vector adjustment data is used to perform potential field diffusion simulation on the initial semantic vector field data, and vector adjustment value derivation is performed to generate non-trajectory vector adjustment data;

[0052] Step S3135: dynamically adjusting the vector field of the initial semantic vector field data by using the semantic trajectory vector adjustment data and the non-trajectory vector adjustment data to generate dynamic semantic vector field data;

[0053] Step S314: performing vector field smoothing optimization according to the dynamic semantic vector field data to generate speech semantic vector field data.

[0054] The present invention can organize and represent semantic information more systematically by dividing the semantic embedding space into multiple grids, and can capture semantic changes at a fine granularity. By calculating the potential energy value, the constructed initial semantic vector field can effectively reflect the distribution characteristics of semantic information in space, identify semantic high and low points, and help determine the priority of semantic information. The initial semantic vector field data is dynamically adjusted by the speech semantic trajectory data, which can reflect the changes of speech signals in real time and enhance the adaptability to speech content. By mapping the speech semantic trajectory to the grid semantic space, the path of semantic flow is clarified. The grid semantic space data is used to grid-remove the semantic trajectory grid sequence data, effectively distinguishing the areas affected by noise or irrelevant signals. Through semantic trajectory vector adjustment and non-trajectory vector adjustment, the initial semantic vector field is accurately optimized. The former performs direct vector adjustment on the main semantic trajectory to ensure the accurate expression of semantic information; the latter performs reasonable smoothing and adjustment on the non-semantic area through potential field diffusion simulation and vector adjustment value derivation, effectively suppresses the interference of noise on the vector field, and maintains the overall coherence of the vector field. The smoothed speech semantic vector field data can more accurately reflect the true characteristics of the speech signal and eliminate the fluctuations caused by unnecessary noise.

[0055] Preferably, step S4 comprises the following steps:

[0056] Step S41: performing noise suppression processing on the speech semantic vector field data through the semantic noise region data, and performing vector field adaptive denoising processing to generate noise-reduced semantic vector field data;

[0057] Step S42: performing semantic trajectory remapping processing on the noise reduction semantic vector field data using the speech semantic trajectory data to generate semantic trajectory remapping data;

[0058] Step S43: Dynamic potential energy constraints are performed on the speech semantic vector field data through the semantic trajectory remapping data, and trajectory drift potential energy is calculated to obtain semantic trajectory drift potential energy data;

[0059] Step S44: Based on a preset noise potential energy threshold, the speech semantic trajectory data is subjected to key semantic noise frame recognition through the semantic trajectory drift potential energy data, and acoustic feature inverse mapping is performed to generate speech key noise frame data;

[0060] Step S45: Perform noise frame repair processing on the framed speech signal data using the speech key noise frame data to generate noise-reduced speech signal data.

[0061] The present invention performs noise suppression processing on semantic noise area data, and the vector field adaptive denoising processing adaptively adjusts the denoising strategy according to the characteristics of the speech signal and the noise distribution, thereby avoiding the loss of useful information caused by excessive denoising, ensuring the high quality of the denoised semantic vector field data, taking measures directly for the identified noise area, and effectively reducing the interference of noise on the speech signal. Through remapping processing, the semantic vector field data after denoising can be aligned with the actual semantic trajectory to ensure that the semantic information of the speech signal can accurately reflect its dynamic changes. Through dynamic potential energy constraints and trajectory drift potential energy calculation, the slight offset or discontinuity of the semantic trajectory after denoising is quantified. By comparing the semantic trajectory drift potential energy data, the key frames that are severely affected by noise are accurately identified, and the acoustic feature inverse mapping converts the identified noise impact into specific frames in the original speech signal, establishing a direct connection for the actual signal repair operation. The noise frame repair processing performed on the framed speech signal data by the speech key noise frame data directly implements the repair on the noise-contaminated frames, which not only reduces the noise interference, but also retains the original features and semantic integrity of the speech signal as much as possible. The generated noise-reduced speech signal data significantly improves the speech quality while maintaining speech clarity.

[0062] Preferably, step S5 comprises the following steps:

[0063] Step S51: performing time-frequency analysis on the noise reduction speech signal data to obtain noise reduction speech spectrum data; performing frequency domain conversion processing on the speech semantic vector field data to generate semantic spectrum field data;

[0064] Step S52: performing multi-base spectrum decomposition processing according to the noise reduction speech spectrum data to generate multi-base frequency speech spectrum data;

[0065] Step S53: using the semantic spectrum field data to perform semantic harmonic association analysis on the multi-fundamental frequency speech spectrum data to generate semantic harmonic association matrix data;

[0066] Step S54: performing harmonic semantic confidence evaluation according to the semantic harmonic association matrix data, and performing speech enhancement mask processing to generate speech enhancement guidance mask data;

[0067] Step S55: semantically enhancing the noise reduction speech spectrum data through the speech enhancement guide mask data, and performing auditory masking spectrum correction to generate corrected speech spectrum data;

[0068] Step S56: Perform speech signal reconstruction processing according to the modified speech spectrum data to generate enhanced speech signal data.

[0069] The present invention can effectively represent the speech signal in the time and frequency domains through time-frequency analysis, revealing the frequency domain characteristics of the speech. Through multi-base spectrum decomposition, the complex speech signal can be decomposed into multiple frequency components, so that the characteristics of each frequency component are more obvious, revealing the subtle differences of each frequency component in the speech signal. By analyzing the harmonic relationship between the semantic spectrum field and the speech spectrum, the important frequency components in the speech signal and their relevance to the semantics can be identified. The harmonic semantic confidence evaluation based on the semantic harmonic association matrix data evaluates the contribution of each spectrum component to conveying semantic information, and the speech enhancement guide mask data generated by the speech enhancement mask processing effectively indicates which spectrum components should be strengthened or weakened. The semantic enhancement processing of the noise-reduced speech spectrum data by the speech enhancement guide mask data specifically enhances the spectrum components carrying important semantic information in the speech signal, combines the auditory masking spectrum correction, and further adjusts the spectrum according to the characteristics of the human auditory system. The generated corrected speech spectrum data improves the speech clarity while making the human ear perceive the speech signal more natural and clear, ensuring that the enhanced speech signal is more in line with the auditory characteristics.

[0070] Preferably, the present invention further provides a speech signal denoising system, which executes the speech signal denoising method as described above, and the speech signal denoising system comprises:

[0071] The adaptive framing module is used to perform adaptive framing processing on the voice signal data to be processed to obtain framed voice signal data; perform fundamental frequency voiceprint feature fusion according to the framed voice signal data to generate initial voiceprint base data;

[0072] The speech semantic analysis module is used to perform acoustic context enhancement processing based on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform time-series semantic trajectory processing based on the semantic embedding space data to obtain speech semantic trajectory data;

[0073] The speech noise recognition module is used to construct a semantic vector field based on the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; mark the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data;

[0074] The speech noise repair module is used to perform noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; perform key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; perform noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data;

[0075] The speech signal enhancement module is used to perform multi-base spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; perform auditory masking spectrum correction based on the multi-base frequency speech spectrum data, and perform speech signal reconstruction processing to generate enhanced speech signal data.

[0076] Preferably, the present invention further provides a speech signal noise reduction device, comprising:

[0077] Memory for storing computer programs;

[0078] A processor, configured to implement the steps of the method for reducing noise of a speech signal as claimed in any one of claims 1 to 8 when executing the computer program. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 A schematic diagram of the steps of the speech signal noise reduction method of the present invention;

[0080] Figure 2 for Figure 1 Detailed implementation steps of step S1 in FIG.

[0081] Figure 3 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0082] Figure 4 for Figure 1 Detailed implementation steps of step S4 in FIG.

[0083] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0084] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0085] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0086] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0087] To achieve this, please refer to Figures 1 to 4 The present invention provides a method for reducing noise of a speech signal, comprising the following steps:

[0088] Step S1: Adaptively frame the speech signal data to be processed to obtain framed speech signal data; perform fundamental frequency voiceprint feature fusion according to the framed speech signal data to generate initial voiceprint base data;

[0089] Step S2: Perform acoustic context enhancement processing on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform temporal semantic trajectory processing on the semantic embedding space data to obtain speech semantic trajectory data;

[0090] Step S3: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; marking the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data;

[0091] Step S4: performing noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; performing key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; performing noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data;

[0092] Step S5: performing multi-base frequency spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; performing auditory masking spectrum correction based on the multi-base frequency speech spectrum data, and performing speech signal reconstruction processing to generate enhanced speech signal data.

[0093] In the embodiment of the present invention, reference Figure 1 The above is a schematic diagram of the steps of the speech signal noise reduction method of the present invention. In this embodiment, the speech signal noise reduction method includes the following steps:

[0094] Step S1: Adaptively frame the speech signal data to be processed to obtain framed speech signal data; perform fundamental frequency voiceprint feature fusion according to the framed speech signal data to generate initial voiceprint base data;

[0095] In an embodiment of the present invention, the voice signal data to be processed is adaptively framed, and the Hamming window can be used to frame the voice signal, with the frame length set to 25ms and the frame shift set to 10ms to ensure a certain overlap between adjacent frames, thereby retaining the continuity of the voice signal. Then, the fundamental frequency voiceprint feature fusion is performed based on the framed voice signal data. First, the Praat software is used to extract the fundamental frequency, resonance peak and other acoustic parameters of each frame of the voice signal to construct a voiceprint feature vector. Next, the dynamic time warping (DTW) algorithm is used to calculate the distance between the voiceprint feature vectors of adjacent frames, and the adjacent frames are weighted fused according to the distance to generate the initial voiceprint base data.

[0096] Step S2: Perform acoustic context enhancement processing on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform temporal semantic trajectory processing on the semantic embedding space data to obtain speech semantic trajectory data;

[0097] In an embodiment of the present invention, based on the acoustic feature vector data, a preset speech recognition model is used to perform frame-by-frame speech recognition on the framed speech signal. A speech recognition system based on a hybrid acoustic model and a language model, such as a speech recognition system based on a hidden Markov model (HMM) and an N-gram language model, or an end-to-end speech recognition system based on a deep neural network (DNN) can be used to obtain frame-level speech text data. Semantic embedding space processing is performed according to the frame-level speech text data. A pre-trained word vector model, such as Word2Vec, GloVe, or BERT, can be used to convert each word or phoneme in the frame-level speech text data into a corresponding word vector to generate semantic embedding space data. Semantic embedding space processing is performed according to the frame-level speech text data. A pre-trained word vector model, such as Word2Vec, GloVe, or BERT, can be used to convert each word or phoneme in the frame-level speech text data into a corresponding word vector to generate semantic embedding space data.

[0098] Step S3: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; marking the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data;

[0099] In an embodiment of the present invention, a self-organizing map (SOM) network can be used to construct a semantic vector field based on speech semantic trajectory data and semantic embedding space data. The speech semantic trajectory data and semantic embedding space data are used as inputs of the SOM network, and through the self-organizing learning process of the network, high-dimensional semantic information is mapped to a two-dimensional plane space to generate speech semantic vector field data. The speech semantic vector field data is analyzed using the characteristics of the speech semantic trajectory data, such as calculating the curl and divergence of the semantic vector field, identifying areas where semantic information has undergone mutations or discontinuities, marking them as potential speech noise areas, and generating semantic noise area data.

[0100] Step S4: performing noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; performing key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; performing noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data;

[0101] In an embodiment of the present invention, the semantic noise area data is used to perform noise suppression processing on the speech semantic vector field data. A masking method can be used to suppress or set the vector value of the semantic noise area to zero to generate noise-reduced semantic vector field data. The speech semantic trajectory data is analyzed using the noise-reduced semantic vector field data to identify key semantic noise frames affected by noise. For example, the difference in semantic vectors at corresponding positions of the semantic trajectory before and after noise reduction can be calculated, and the frames with larger differences can be marked as key semantic noise frames to generate speech key noise frame data. Noise frame repair processing is performed on the framed speech signal data according to the speech key noise frame data. A speech generation model based on deep learning, such as WaveNet or WaveRNN, can be used to repair key noise frames, generate estimated clean speech frames, replace the corresponding frames in the original speech signal, and generate noise-reduced speech signal data.

[0102] Step S5: performing multi-base frequency spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; performing auditory masking spectrum correction based on the multi-base frequency speech spectrum data, and performing speech signal reconstruction processing to generate enhanced speech signal data.

[0103] In an embodiment of the present invention, the noise reduction speech signal data is subjected to multi-base spectrum decomposition processing, for example, using the empirical mode decomposition (EMD) algorithm to decompose it into a plurality of intrinsic mode functions (IMFs) of different frequency scales, and generate multi-baseband speech spectrum data. Auditory masking spectrum correction is performed based on the multi-baseband speech spectrum data. Based on the auditory masking effect of the human ear, the multi-baseband speech spectrum data is corrected, for example, the energy of the noise below the masking threshold is reduced to improve the intelligibility of the speech. The corrected multi-baseband speech spectrum data is subjected to speech signal reconstruction processing. The various IMF components are superimposed to generate enhanced speech signal data, and the speech signal noise reduction is completed.

[0104] Preferably, step S1 comprises the following steps:

[0105] Step S11: performing voice endpoint detection on the voice signal data to be processed and performing silent segment elimination processing to generate initial voice signal data;

[0106] Step S12: extracting the fundamental frequency period according to the initial speech signal data to obtain fundamental frequency contour feature data;

[0107] Step S13: performing adaptive framing processing according to the initial voice signal data to obtain framed voice signal data;

[0108] Step S14: performing fast Fourier transform processing on the framed speech signal data, and calculating the Mel frequency cepstrum coefficients according to the Mel filter to generate Mel frequency spectrum feature matrix data;

[0109] Step S15: performing dynamic time warping processing on the fundamental frequency profile feature data through the Mel frequency spectrum feature matrix data, and performing fundamental frequency voiceprint feature fusion to obtain fundamental frequency fusion feature data;

[0110] Step S16: Estimating the voiceprint distribution probability based on the baseband fusion feature data, and performing Gaussian fuzzy generalization processing to generate initial voiceprint base data.

[0111] As an example of the present invention, refer to Figure 2 As shown, Figure 1 Detailed implementation steps of step S1 in the flowchart, in this example, step S1 includes:

[0112] Step S11: performing voice endpoint detection on the voice signal data to be processed and performing silent segment elimination processing to generate initial voice signal data;

[0113] In an embodiment of the present invention, a double threshold method based on energy and zero-crossing rate can be used to detect speech endpoints of input original speech signal data. First, an energy threshold and a zero-crossing rate threshold are set, and the short-time energy and short-time zero-crossing rate of the speech signal are calculated. When the short-time energy and the short-time zero-crossing rate both exceed the preset threshold, it is determined as the starting point of the speech segment; when the short-time energy and the short-time zero-crossing rate are both lower than the preset threshold, it is determined as the end point of the speech segment. According to the detected speech endpoints, the silent segments are eliminated, and the segments containing speech information are retained to generate the initial speech signal data.

[0114] Step S12: extracting the fundamental frequency period according to the initial speech signal data to obtain fundamental frequency contour feature data;

[0115] In the embodiment of the present invention, the initial speech signal data is preprocessed, including pre-emphasis and windowing operations, to highlight the high-frequency information and smooth signal in the speech signal. Then, the autocorrelation function of each frame of the speech signal is calculated, and the first peak position of the autocorrelation function is found, and the peak corresponds to the fundamental frequency period. In order to improve the accuracy of the fundamental frequency period extraction, a dynamic threshold method can be used to determine the peak position, and the extracted fundamental frequency period is subjected to median filtering to remove abnormal values, and finally obtain smooth fundamental frequency contour feature data.

[0116] Step S13: performing adaptive framing processing according to the initial voice signal data to obtain framed voice signal data;

[0117] In an embodiment of the present invention, the initial speech signal data is subjected to adaptive framing processing. The speech signal can be framed using a Hamming window, with the frame length set to 25ms and the frame shift set to 10ms to ensure a certain overlap between adjacent frames, thereby retaining the continuity of the speech signal and obtaining framed speech signal data. For example, a continuous speech signal is divided into several frames using a short-time Fourier transform (STFT) technique. The length and overlap of each frame can be adjusted according to the characteristics of the speech signal, and a frame length of 20ms to 30ms and an overlap of 50% are usually selected. In this way, framed speech signal data can be generated.

[0118] Step S14: performing fast Fourier transform processing on the framed speech signal data, and calculating the Mel frequency cepstrum coefficients according to the Mel filter to generate Mel frequency spectrum feature matrix data;

[0119] In the embodiment of the present invention, the framed speech signal data is processed by fast Fourier transform (FFT), the time domain signal is converted to the frequency domain, and the spectrum information of the speech signal is obtained. Then, the spectrum is filtered according to the Mel filter bank to simulate the auditory characteristics of the human ear and obtain the energy distribution of different frequency bands. Finally, the filtered energy spectrum is logarithmically operated and discrete cosine transformed (DCT) to obtain the Mel frequency cepstrum coefficient (MFCC), and the MFCC coefficients of all frames constitute the Mel spectrum feature matrix data.

[0120] Step S15: performing dynamic time warping processing on the fundamental frequency profile feature data through the Mel frequency spectrum feature matrix data, and performing fundamental frequency voiceprint feature fusion to obtain fundamental frequency fusion feature data;

[0121] In an embodiment of the present invention, a dynamic time warping (DTW) algorithm is used to align the Mel spectrum feature matrix data and the fundamental frequency contour feature data to eliminate the differences between different speech signals on the time axis. The DTW algorithm achieves time warping by finding the best matching path between two time series to minimize the cumulative distance. Based on the aligned feature sequence, fundamental frequency voiceprint feature fusion is performed. A linear weighted fusion method can be used to set the weight coefficient based on experience or experimental results, and the Mel spectrum features and fundamental frequency contour features are weighted and summed to obtain fundamental frequency fusion feature data.

[0122] Step S16: Estimating the voiceprint distribution probability based on the baseband fusion feature data, and performing Gaussian fuzzy generalization processing to generate initial voiceprint base data.

[0123] In an embodiment of the present invention, a Gaussian mixture model (GMM) can be used to estimate the probability of voiceprint distribution based on the baseband fusion feature data. GMM can model complex speech feature distribution as a weighted sum of multiple Gaussian distributions, each of which represents a voiceprint pattern. By estimating the parameters of GMM, the probability that each speech frame belongs to a different voiceprint pattern can be obtained. Through this probability estimation, the relationship between different voiceprint data can be better understood. Subsequently, the voiceprint distribution is subjected to Gaussian fuzzy generalization processing to reduce the sensitivity of the feature data and the influence of noise, thereby generating initial voiceprint base data.

[0124] Preferably, step S2 comprises the following steps:

[0125] Step S21: performing acoustic context enhancement processing according to the initial voiceprint base data to generate acoustic feature vector data;

[0126] Step S22: performing frame-by-frame speech recognition on the framed speech signal data using a preset speech recognition model based on the acoustic feature vector data to obtain frame-level speech text data;

[0127] Step S23: performing semantic embedding space processing according to the frame-level speech text data to obtain semantic embedding space data;

[0128] Step S24: semantic space mapping is performed on the framed speech signal data by using the semantic embedding space data to generate framed speech semantic coordinate data;

[0129] Step S25: performing time-series semantic trajectory processing according to the framed speech semantic coordinate data to obtain speech semantic trajectory data.

[0130] As an example of the present invention, refer to Figure 3 As shown, Figure 1 Detailed implementation steps of step S2 in the embodiment are shown in the flowchart. In this embodiment, step S2 includes:

[0131] Step S21: performing acoustic context enhancement processing according to the initial voiceprint base data to generate acoustic feature vector data;

[0132] In an embodiment of the present invention, a recurrent neural network (RNN), such as a long short-term memory network (LSTM), is used to learn the temporal characteristics of speech signals. The initial voiceprint base data is used as the input of the RNN. Through the network's cyclic structure and nonlinear activation function, the contextual information of the speech signal can be captured, thereby generating more representative acoustic feature vector data. For example, the initial voiceprint base data of each frame of speech and the voiceprint data of the previous and next frames can be spliced ​​into a vector as the input of the LSTM. The LSTM network will learn the acoustic feature relationship between adjacent frames and output an acoustic feature vector containing richer contextual information.

[0133] Step S22: performing frame-by-frame speech recognition on the framed speech signal data using a preset speech recognition model based on the acoustic feature vector data to obtain frame-level speech text data;

[0134] In an embodiment of the present invention, based on the acoustic feature vector data, a preset speech recognition model is used to perform frame-by-frame speech recognition on the framed speech signal data. A speech recognition system using a hybrid acoustic model and a language model can be selected, such as a speech recognition system based on a hidden Markov model (HMM) and an N-gram language model, or an end-to-end speech recognition system based on a deep neural network (DNN). The acoustic feature vector of each frame is input into the speech recognition model, and the speech recognition result of the corresponding frame is obtained by decoding, and finally the frame-level speech text data, that is, the phoneme or word sequence corresponding to each speech frame, is obtained.

[0135] Step S23: performing semantic embedding space processing according to the frame-level speech text data to obtain semantic embedding space data;

[0136] In an embodiment of the present invention, semantic embedding space processing is performed according to the frame-level speech text data, and a pre-trained word vector model, such as Word2Vec, GloVe or BERT, can be used. These models can map each word or phoneme to a high-dimensional vector space, so that semantically similar words or phonemes are closer in the vector space. Each word or phoneme in the frame-level speech text data is converted into a corresponding word vector, and the word vector of each speech frame is averaged or weighted averaged to obtain a semantic embedding vector for each speech frame, and finally constitute the semantic embedding space data.

[0137] Step S24: semantic space mapping is performed on the framed speech signal data by using the semantic embedding space data to generate framed speech semantic coordinate data;

[0138] In an embodiment of the present invention, semantic space mapping is performed on the framed speech signal data through semantic embedding space data. The acoustic feature vector and semantic embedding vector of each speech frame can be spliced ​​into a new vector, and mapped to a new semantic space using linear transformation or nonlinear transformation. For example, a multi-layer perceptron (MLP) can be used to perform nonlinear transformation on the spliced ​​vector, learn the complex relationship between acoustic features and semantic features, and map each speech frame to a point in the semantic space, whose coordinates are the framed speech semantic coordinate data, to achieve the conversion of the speech signal from the acoustic space to the semantic space.

[0139] Step S25: performing time-series semantic trajectory processing according to the framed speech semantic coordinate data to obtain speech semantic trajectory data.

[0140] In the embodiment of the present invention, the time-series semantic trajectory processing is performed according to the framed speech semantic coordinate data, and the coordinates of each speech frame in the semantic space can be connected to form a continuous trajectory, i.e., speech semantic trajectory data. This trajectory reflects the changing trend of the semantic information of the speech signal over time. In order to better capture the dynamic information of the semantic trajectory, the trajectory can be smoothed, for example, using a moving average filter or a Gaussian filter to remove high-frequency noise and obtain a smoother speech semantic trajectory.

[0141] Preferably, step S23 includes the following steps:

[0142] Step S231: converting the frame-level speech text data into a high-dimensional semantic vector to generate semantic original vector data;

[0143] Step S232: constructing a high-dimensional manifold structure according to the semantic original vector data to obtain semantic manifold structure data;

[0144] Step S233: using the frame-level speech text data to identify the core semantic factors of the semantic manifold structure data, and generating core semantic factor data;

[0145] Step S234: using a preset speech corpus to perform contextual semantic embedding matching on the frame-level speech text data, and generating semantic corpus through a generative adversarial network to obtain diversified corpus data;

[0146] Step S235: decoupling semantic features according to the diversified corpus data, and performing semantic association network processing to generate semantic relationship network data;

[0147] Step S236: Dynamically semantically fuse the semantic manifold structure data through the semantic relationship network data to obtain dynamic semantic embedding data;

[0148] Step S237: semantically calibrate the dynamic semantic embedding data, and construct a KD tree fast index based on the core semantic factor data, so as to obtain semantic embedding space data.

[0149] In an embodiment of the present invention, a pre-trained language model, such as BERT or GPT, is used. The frame-level speech text data is input into the pre-trained language model to obtain the semantic vector representation corresponding to each word or subword. In order to better capture the semantic information at the sentence level, the vector representation corresponding to the [CLS] tag in the BERT model can be used, or the vector representation of all words in the sentence can be averaged and pooled to generate semantic original vector data. According to the semantic original vector data, a high-dimensional manifold structure is constructed, and a dimensionality reduction algorithm such as t-SNE can be used. The t-SNE algorithm can map high-dimensional data to a low-dimensional space while retaining the distance relationship between data points, thereby revealing the inherent manifold structure of the data. The semantic original vector data is used as input, mapped to a two-dimensional or three-dimensional space by the t-SNE algorithm, and a Delaunay triangulation or a k-nearest neighbor graph is constructed to obtain semantic manifold structure data. The core semantic factors of the semantic manifold structure data are identified by frame-level speech text data, and an algorithm based on graph theory can be used. First, each word in the frame-level speech text data is mapped to a corresponding node in the semantic manifold structure. Then, the degree centrality, betweenness centrality or PageRank value of each node are calculated to measure the importance of the node in the graph. Finally, the nodes with the highest centrality index are selected as core semantic factors and marked on the semantic manifold structure to generate core semantic factor data. The frame-level speech text data is matched with contextual semantic embedding using the preset speech corpus, and the semantic similarity at the sentence level can be calculated using models such as SentenceBERT. The frame-level speech text data is used as a query sentence to retrieve semantically similar sentences in the preset speech corpus. Then, the adversarial generative network (GAN) is used for semantic corpus generation. The retrieved similar sentences are used as training data for GAN to train a generator network so that it can generate new sentences that are semantically similar to the input sentences, thereby expanding the diversity of the corpus and obtaining diversified corpus data. The MINE method is used to estimate the mutual information between different semantic features, and feature combinations with low mutual information are selected for decoupling, thereby eliminating redundant information between semantic features. Then, methods such as graph attention network (GAT) or graph convolutional network (GCN) are used for semantic association network processing. The words in the diverse corpus data are used as nodes, and the co-occurrence relationship or semantic similarity relationship between words is used as the edge to construct a semantic association network, and the hidden representation of the node is learned using GAT or GCN to capture the semantic association relationship between words and generate semantic relationship network data. The graph neural network (GNN) model can be used to dynamically semantically fuse the semantic manifold structure data through the semantic relationship network data. The semantic manifold structure data and the semantic relationship network data are used as the input of the GNN model, and the local structure information of the node and the global semantic association information are fused through the graph convolution or message passing mechanism, and the node representation is updated to obtain dynamic semantic embedding data.The dynamic semantic embedding data can be semantically calibrated using methods such as Procrustes analysis. The dynamic semantic embedding data is aligned with a predefined semantic space, such as WordNet or ConceptNet, to provide a semantic interpretation for the dynamic semantic embedding data. Then, a KD tree fast index is constructed based on the core semantic factor data. The core semantic factors are used as nodes of the KD tree, and hierarchical division is performed according to their coordinates in the semantic embedding space. The KD tree index structure is constructed for fast searching of the nearest neighbor nodes, and finally the semantic embedding space data is obtained.

[0150] Preferably, step S3 comprises the following steps:

[0151] Step S31: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data;

[0152] Step S32: performing grid tensor feature analysis on the speech semantic vector field data to generate grid semantic tensor data; performing semantic vector streamline tracing on the speech semantic vector field data through the grid semantic tensor data to obtain semantic vector streamline data;

[0153] Step S33: performing streamline stability evaluation according to the semantic vector streamline data to generate streamline stability evaluation data;

[0154] Step S34: performing semantic eddy current detection on the speech semantic vector field data based on the semantic vector streamline data to generate speech eddy current area data;

[0155] Step S35: calculating the average speed of the semantic vector streamline data through a preset multi-scale sliding window to generate semantic streamline speed sequence data;

[0156] Step S36: extracting semantic acceleration from speech eddy current region data based on semantic streamline velocity sequence data, and quantifying eddy current intensity to generate eddy current intensity data;

[0157] Step S37: Mark the speech noise area of ​​the speech semantic vector field data using the eddy current intensity data, the speech eddy current area data, and the streamline stability evaluation data to generate semantic noise area data.

[0158] In an embodiment of the present invention, an interpolation method can be used to construct a semantic vector field based on speech semantic trajectory data and semantic embedding space data. First, the semantic embedding space is regarded as a continuous field, and each point in the speech semantic trajectory data corresponds to a position in the semantic space. Then, the discrete trajectory points are expanded into a continuous vector field using methods such as radial basis function (RBF) interpolation or inverse distance weighted interpolation. For any point in the semantic space, the semantic vector of the point can be calculated based on its distance from the trajectory point and the semantic vector corresponding to the trajectory point, thereby obtaining complete speech semantic vector field data for characterizing the change trend of the speech signal at different semantic positions. The semantic space where the speech semantic vector field is located is divided into a uniform grid, and the gradient tensor of the vector field is calculated at each grid point. The gradient tensor can describe the direction and rate of change of the vector field at the point, reflecting the local change characteristics of the semantic information. The Runge-Kutta method can be used to track the semantic vector streamlines of the speech semantic vector field data using the grid semantic tensor data. Starting from the starting point of the semantic vector field, the position of the next point is iteratively calculated along the gradient direction until the end point of the vector field or the preset number of steps is reached, thereby obtaining a complete semantic vector streamline. Repeating the above steps, multiple semantic vector streamlines can be obtained to form semantic vector streamline data. The Lyapunov index can be used. The Lyapunov index can measure the divergence speed of adjacent trajectories of the system in the phase space and is used to judge the degree of chaos of the system. For each semantic vector streamline, its Lyapunov index can be calculated, and the stability of the streamline can be judged according to the positive or negative nature of the index. A positive Lyapunov index indicates that the streamline is unstable and easily disturbed by noise; a negative Lyapunov index indicates that the streamline is stable and can resist noise interference. Curl is a vector field that describes the degree of rotation of a vector field at a certain point. For the speech semantic vector field, the curl of each point can be calculated, and the presence of a semantic vortex can be judged according to the size and direction of the curl. For example, a curl threshold can be set. When the curl modulus of a point is greater than the threshold, it is considered that a semantic vortex exists at the point. Set multiple sliding windows of different sizes, such as 5 points, 10 points, and 20 points. Then, slide the sliding window along each semantic vector streamline, and calculate the average speed of all points in the window. Specifically, each point can be regarded as a time step, and the speed of the point can be obtained by dividing the difference between the semantic vectors at two adjacent time steps by the time interval. Set multiple sliding windows of different sizes, such as 5 points, 10 points, and 20 points. Then, slide the sliding window along each semantic vector streamline, and calculate the average speed of all points in the window. Specifically, each point can be regarded as a time step, and the speed of the point can be obtained by dividing the difference between the semantic vectors at two adjacent time steps by the time interval. Set multiple sliding windows of different sizes, such as 5 points, 10 points, and 20 points.Then, the sliding window is slid along each semantic vector streamline to calculate the average speed of all points in the window. Specifically, each point can be regarded as a time step, and the speed of the point can be obtained by dividing the difference between the semantic vectors of two adjacent time steps by the time interval.

[0159] Preferably, step S31 includes the following steps:

[0160] Step S311: performing semantic space gridding processing on the semantic embedding space data to generate grid semantic space data; performing initial semantic potential energy value calculation on the grid semantic space data to generate initial semantic potential energy data;

[0161] Step S312: constructing an initial vector field for the grid semantic space data using the initial semantic potential energy data to obtain initial semantic vector field data;

[0162] Step S313: dynamically adjusting the initial semantic vector field data using the speech semantic trajectory data to generate dynamic semantic vector field data; wherein step S313 is specifically as follows:

[0163] Step S3131: using the speech semantic trajectory data to perform semantic trajectory mapping processing on the grid semantic space data to generate semantic trajectory grid sequence data;

[0164] Step S3132: using the grid semantic space data to perform grid elimination on the semantic trajectory grid sequence data to obtain non-semantic trajectory grid data;

[0165] Step S3133: adjusting the semantic trajectory vector of the initial semantic vector field data using the speech semantic trajectory data based on the semantic trajectory grid sequence data to generate semantic trajectory vector adjustment data;

[0166] Step S3134: Based on the non-semantic trajectory grid data, the semantic trajectory vector adjustment data is used to perform potential field diffusion simulation on the initial semantic vector field data, and vector adjustment value derivation is performed to generate non-trajectory vector adjustment data;

[0167] Step S3135: dynamically adjusting the vector field of the initial semantic vector field data by using the semantic trajectory vector adjustment data and the non-trajectory vector adjustment data to generate dynamic semantic vector field data;

[0168] Step S314: performing vector field smoothing optimization according to the dynamic semantic vector field data to generate speech semantic vector field data.

[0169] In an embodiment of the present invention, the semantic embedding space is regarded as a multidimensional space, and each dimension is equally divided into several intervals according to a preset grid size, so that the entire semantic space is divided into multiple hypercube grids of equal size. Each grid represents a local area in the semantic space, and grid semantic space data is generated. Then, the initial semantic potential energy value is calculated for the grid semantic space data. The initial semantic potential energy value can be represented by the semantic vector density or information entropy in each grid. The semantic embedding space is regarded as a multidimensional space, and each dimension is equally divided into several intervals according to a preset grid size, so that the entire semantic space is divided into multiple hypercube grids of equal size. Each grid represents a local area in the semantic space, and grid semantic space data is generated. Then, the initial semantic potential energy value is calculated for the grid semantic space data, and the initial semantic potential energy value can be represented by the semantic vector density or information entropy in each grid. The speech semantic trajectory data is used to perform semantic trajectory mapping processing on the grid semantic space data, and the nearest neighbor search method can be used. For each point in the speech semantic trajectory data, find the grid closest to it in the semantic space, and map the point to the grid to generate semantic trajectory grid sequence data. The semantic trajectory grid sequence data is meshed using the grid semantic space data, and set operations can be used. All semantic trajectory grids are removed from the grid semantic space data to obtain non-semantic trajectory grid data, which is used to identify the grid area that the speech semantic trajectory has not passed through. For each semantic trajectory grid, according to its position and direction in the speech semantic trajectory, the direction and size of the initial semantic vector on the grid are adjusted. For example, the direction of the initial semantic vector can be adjusted to be consistent with the tangent direction of the speech semantic trajectory at the grid, and the size of the initial semantic vector can be set to be proportional to the speed of the speech semantic trajectory at the grid, and the semantic trajectory vector adjustment data is generated. The heat conduction equation can be used for simulation, for example, the semantic trajectory vector adjustment data is regarded as a heat source, the non-semantic trajectory grid is regarded as a heat conduction medium, and the heat conduction equation is used to simulate the process of heat diffusion from the heat source to the surrounding. According to the simulation results, the temperature change on each non-semantic trajectory grid can be obtained, and it is converted into a corresponding vector adjustment value to generate non-trajectory vector adjustment data. The initial semantic vector field data is dynamically adjusted by using the semantic trajectory vector adjustment data and the non-trajectory vector adjustment data, and a weighted average method can be used. The initial semantic vector, the semantic trajectory vector adjustment amount, and the non-trajectory vector adjustment amount are weighted averaged to obtain the final semantic vector on each grid and generate dynamic semantic vector field data.

[0170] Preferably, step S4 comprises the following steps:

[0171] Step S41: performing noise suppression processing on the speech semantic vector field data through the semantic noise region data, and performing vector field adaptive denoising processing to generate noise-reduced semantic vector field data;

[0172] Step S42: performing semantic trajectory remapping processing on the noise reduction semantic vector field data using the speech semantic trajectory data to generate semantic trajectory remapping data;

[0173] Step S43: Dynamic potential energy constraints are performed on the speech semantic vector field data through the semantic trajectory remapping data, and trajectory drift potential energy is calculated to obtain semantic trajectory drift potential energy data;

[0174] Step S44: Based on a preset noise potential energy threshold, the speech semantic trajectory data is subjected to key semantic noise frame recognition through the semantic trajectory drift potential energy data, and acoustic feature inverse mapping is performed to generate speech key noise frame data;

[0175] Step S45: Perform noise frame repair processing on the framed speech signal data using the speech key noise frame data to generate noise-reduced speech signal data.

[0176] As an example of the present invention, refer to Figure 4 As shown, Figure 1 Detailed implementation steps of step S4 in the embodiment are shown in the flowchart. In this embodiment, step S4 includes:

[0177] Step S41: performing noise suppression processing on the speech semantic vector field data through the semantic noise region data, and performing vector field adaptive denoising processing to generate noise-reduced semantic vector field data;

[0178] In an embodiment of the present invention, a masking method can be used to perform noise suppression processing on speech semantic vector field data through semantic noise area data. First, a binary mask is generated according to the semantic noise area data, and the grid points corresponding to the semantic noise area are marked as 0, and other areas are marked as 1. Then, the mask is multiplied with the speech semantic vector field data to suppress the vector value of the noise area and retain the vector information of the non-noise area. In order to avoid excessive suppression, a soft masking method can be used to multiply the vector value of the noise area by a weight less than 1 instead of directly setting it to zero. The anisotropic diffusion equation can be used to control the direction and speed of diffusion, for example, diffusing along the direction of the semantic vector and slowing down the diffusion speed at the semantic boundary to avoid excessive smoothing of the semantic information, and finally generating noise-reduced semantic vector field data.

[0179] Step S42: performing semantic trajectory remapping processing on the noise reduction semantic vector field data using the speech semantic trajectory data to generate semantic trajectory remapping data;

[0180] In an embodiment of the present invention, the nearest neighbor interpolation method is used for semantic trajectory remapping. First, the semantic vector field after noise reduction is regarded as a new semantic space, and each grid point in the space corresponds to a semantic vector. Then, for each trajectory point in the speech semantic trajectory data, the distance between it and all grid points in the new semantic space is calculated. Common distance measurement methods such as Euclidean distance and cosine distance can be selected to calculate the distance. Find the nearest grid point and map the trajectory point to the grid point. In this way, the mapping of the original speech semantic trajectory to the noise reduction semantic vector field is completed. Assuming that there is a trajectory point A in the original semantic space, its coordinates are (1,2), and in the semantic space after noise reduction, the coordinates of the grid point closest to it are (1.1,1.9), then the trajectory point A is remapped to the grid point with coordinates (1.1,1.9). Perform the same operation on all trajectory points in the speech semantic trajectory data to obtain the semantic trajectory remapping data.

[0181] Step S43: Dynamic potential energy constraints are performed on the speech semantic vector field data through the semantic trajectory remapping data, and trajectory drift potential energy is calculated to obtain semantic trajectory drift potential energy data;

[0182] In an embodiment of the present invention, a Gaussian function is constructed with each remapped trajectory point as the center, and the value of the function is the largest at the trajectory point and decays as the distance from the trajectory point increases. The amplitude of the Gaussian function can be adjusted according to the importance of the trajectory point. For example, the speed or curvature of the trajectory point can be used as a weight. By superimposing the Gaussian functions corresponding to all trajectory points, a potential constraint field that describes the influence of the semantic trajectory can be obtained. With the trajectory point as the center, the closer the area is to the trajectory point, the stronger its potential constraint effect. Then, the trajectory drift potential energy calculation is performed. For each point on the remapped semantic trajectory, the potential energy difference between the corresponding positions on the original semantic vector field and the denoised semantic vector field is calculated, and the difference is used as the trajectory drift potential energy of the point, which reflects the degree of influence of the noise suppression processing on the speech semantic trajectory, and the semantic trajectory drift potential energy data is obtained.

[0183] Step S44: Based on a preset noise potential energy threshold, the speech semantic trajectory data is subjected to key semantic noise frame recognition through the semantic trajectory drift potential energy data, and acoustic feature inverse mapping is performed to generate speech key noise frame data;

[0184] In an embodiment of the present invention, the speech semantic trajectory data is used to identify key semantic noise frames. If the drift potential energy of the semantic trajectory point corresponding to a certain speech frame is higher than a preset threshold, it is considered that the speech frame is interfered by noise and is marked as a key semantic noise frame. Because a larger drift potential energy indicates that the position of the speech frame in the semantic space has changed significantly, which is usually caused by noise. Then, the acoustic features of the key semantic noise frame are inversely mapped. Since the previous processing steps are all performed in the semantic space, it is necessary to map the key semantic noise frame back to the original acoustic feature space. A pre-trained acoustic model or a deep neural network can be used to convert semantic features into acoustic features to generate speech key noise frame data.

[0185] Step S45: Perform noise frame repair processing on the framed speech signal data using the speech key noise frame data to generate noise-reduced speech signal data.

[0186] In an embodiment of the present invention, it is ensured that both the framed speech signal data and the speech key noise frame data are prepared. The framed speech signal data is usually a two-dimensional array, in which each row represents a time frame and each column represents a feature dimension. The speech key noise frame data records the frame index identified as noise. The acoustic features of the speech key noise frame and the acoustic features of the clean speech frames before and after it are taken as input, and the acoustic feature vector of the key noise frame is adjusted in combination with the speech semantic vector field data to make it closer to the acoustic feature vector of the target speech signal. The WaveNet model is trained to learn the mapping relationship from noisy speech to clean speech. After the training is completed, the model can be used to repair the key noise frame to generate an estimated clean speech frame.

[0187] Preferably, step S5 comprises the following steps:

[0188] Step S51: performing time-frequency analysis on the noise reduction speech signal data to obtain noise reduction speech spectrum data; performing frequency domain conversion processing on the speech semantic vector field data to generate semantic spectrum field data;

[0189] Step S52: performing multi-base spectrum decomposition processing according to the noise reduction speech spectrum data to generate multi-base frequency speech spectrum data;

[0190] Step S53: using the semantic spectrum field data to perform semantic harmonic association analysis on the multi-fundamental frequency speech spectrum data to generate semantic harmonic association matrix data;

[0191] Step S54: performing harmonic semantic confidence evaluation according to the semantic harmonic association matrix data, and performing speech enhancement mask processing to generate speech enhancement guidance mask data;

[0192] Step S55: semantically enhancing the noise reduction speech spectrum data through the speech enhancement guide mask data, and performing auditory masking spectrum correction to generate corrected speech spectrum data;

[0193] Step S56: Perform speech signal reconstruction processing according to the modified speech spectrum data to generate enhanced speech signal data.

[0194] In an embodiment of the present invention, a short-time Fourier transform (STFT) can be used to perform time-frequency analysis on the noise reduction speech signal data. The noise reduction speech signal data is divided into a plurality of overlapping short-time frames, and a Fourier transform is performed on each frame to obtain the spectrum information of the frame. The spectrum information of all frames is arranged in chronological order to obtain the noise reduction speech spectrum data. Each semantic vector in the semantic vector field is projected onto a predefined semantic basis vector to obtain the components of the semantic vector in different semantic dimensions. Then, a Fourier transform is performed on the components on each semantic dimension to obtain the representation of the semantic dimension in the frequency domain. The empirical mode decomposition (EMD) algorithm can be used to perform multi-base spectrum decomposition processing according to the noise reduction speech spectrum data. The EMD algorithm can decompose a non-stationary signal into a plurality of intrinsic mode functions (IMFs) of different frequency scales, and each IMF component represents the oscillation mode of the original signal within a specific frequency range. The noise reduction speech spectrum data is used as input, and it is decomposed into a plurality of IMF components by the EMD algorithm to generate multi-baseband speech spectrum data, and the instantaneous frequency of each IMF component is calculated, which represents the frequency value of the IMF component at a specific time point. Then, according to the instantaneous frequency, each IMF component is associated with the semantic information of the corresponding frequency in the semantic spectrum field data. The similarity between the IMF component and the semantic spectrum at a specific frequency point can be calculated, for example, using cosine similarity. Finally, the similarities of all time points and all IMF components are combined to generate semantic harmonic association matrix data. According to the size of the similarity value, a threshold can be set, and the IMF components with a similarity higher than the threshold are marked as strongly associated with the semantic information and given a higher confidence; the IMF components with a similarity lower than the threshold are marked as weakly associated with the semantic information and given a lower confidence. Then, speech enhancement mask processing is performed. According to the harmonic semantic confidence, speech enhancement guide mask data is generated. The area with high confidence corresponds to important speech information, and the mask value is close to 1; the area with low confidence corresponds to noise or unimportant speech information, and the mask value is close to 0. The speech enhancement guide mask data is multiplied with the noise reduction speech spectrum data to enhance the speech components related to important semantic information and suppress the speech components related to noise or unimportant semantic information. Then, the auditory masking spectrum is corrected. According to the auditory masking effect of the human ear, the semantically enhanced speech spectrum is corrected, for example, the energy of the noise below the masking threshold is reduced to improve the intelligibility of the speech, and the corrected speech spectrum data is generated. The inverse short-time Fourier transform (ISTFT) can be used. The corrected spectrum data is subjected to ISTFT to obtain the enhanced speech signal data, completing the speech signal enhancement process.

[0195] Preferably, the present invention further provides a speech signal denoising system, which executes the speech signal denoising method as described above, and the speech signal denoising system comprises:

[0196] The adaptive framing module is used to perform adaptive framing processing on the voice signal data to be processed to obtain framed voice signal data; perform fundamental frequency voiceprint feature fusion according to the framed voice signal data to generate initial voiceprint base data;

[0197] The speech semantic analysis module is used to perform acoustic context enhancement processing based on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform time-series semantic trajectory processing based on the semantic embedding space data to obtain speech semantic trajectory data;

[0198] The speech noise recognition module is used to construct a semantic vector field based on the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; mark the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data;

[0199] The speech noise repair module is used to perform noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; perform key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; perform noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data;

[0200] The speech signal enhancement module is used to perform multi-base spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; perform auditory masking spectrum correction based on the multi-base frequency speech spectrum data, and perform speech signal reconstruction processing to generate enhanced speech signal data.

[0201] Preferably, the present invention further provides a speech signal noise reduction device, comprising:

[0202] Memory for storing computer programs;

[0203] A processor, configured to implement the steps of the method for reducing noise of a speech signal as claimed in any one of claims 1 to 8 when executing the computer program.

[0204] The beneficial effect of the present application is that the processed speech signal data is adaptively framed and the continuous speech signal is divided into several frames, which can fully capture the dynamic characteristics of the speech signal. Compared with the traditional method, the adaptive frame processing can not only provide more accurate time resolution, but also better reflect the instantaneous changes of the signal. The speech signal is converted into text information using speech recognition technology. On this basis, the text information is mapped to the semantic embedding space using a pre-trained semantic model to construct a speech representation containing rich semantic information. The semantic embedding space is further constructed into a semantic vector field. By analyzing the semantic features of the speech signal that change over time, the system can identify the key parts of the speech signal and reduce the problem of speech signal loss caused by silence or background noise interference. The semantic change trajectory of the speech signal in the time dimension is mapped to the vector field. By analyzing the topological structure of the semantic vector field and the motion characteristics of the speech semantic trajectory, such as semantic vortex, trajectory drift, etc., potential noise areas can be effectively identified, especially the weak signal part under the noise background. Targeted noise suppression processing is performed on the semantic vector field, such as suppressing or setting the vectors in the noise area to zero, identifying key speech frames affected by noise, and using deep learning models to perform fine repair on these frames, thereby minimizing noise while retaining the original speech information. The denoised speech is subjected to multi-base spectrum decomposition, and harmonic correlation analysis is performed in combination with semantic information to evaluate the correlation between different frequency components and semantic information, and a speech enhancement guidance mask is generated accordingly. Finally, the speech spectrum is enhanced using the mask and corrected in combination with the auditory masking effect, ultimately reconstructing a clear, highly intelligible enhanced speech signal, significantly improving the user experience.

[0205] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0206] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for reducing noise of a speech signal, characterized in that: The following steps are involved: Step S1: Adaptively frame the speech signal data to be processed to obtain framed speech signal data; Perform fundamental frequency voiceprint feature fusion according to the framed speech signal data to generate initial voiceprint base data; Step S2: Perform acoustic context enhancement processing on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform temporal semantic trajectory processing on the semantic embedding space data to obtain speech semantic trajectory data; Step S3: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; marking the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data; Step S4: performing noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; performing key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; performing noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data; Step S5: performing multi-base frequency spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; Auditory masking spectrum correction is performed based on the multi-baseband speech spectrum data, and speech signal reconstruction processing is performed to generate enhanced speech signal data.

2. The method for reducing noise of a speech signal according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: performing voice endpoint detection on the voice signal data to be processed and performing silent segment elimination processing to generate initial voice signal data; Step S12: extracting the fundamental frequency period according to the initial speech signal data to obtain fundamental frequency contour feature data; Step S13: performing adaptive framing processing according to the initial voice signal data to obtain framed voice signal data; Step S14: performing fast Fourier transform processing on the framed speech signal data, and calculating the Mel frequency cepstrum coefficients according to the Mel filter to generate Mel frequency spectrum feature matrix data; Step S15: performing dynamic time warping processing on the fundamental frequency profile feature data through the Mel frequency spectrum feature matrix data, and performing fundamental frequency voiceprint feature fusion to obtain fundamental frequency fusion feature data; Step S16: Estimating the voiceprint distribution probability based on the baseband fusion feature data, and performing Gaussian fuzzy generalization processing to generate initial voiceprint base data.

3. The method for reducing noise of a speech signal according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing acoustic context enhancement processing according to the initial voiceprint base data to generate acoustic feature vector data; Step S22: performing frame-by-frame speech recognition on the framed speech signal data using a preset speech recognition model based on the acoustic feature vector data to obtain frame-level speech text data; Step S23: performing semantic embedding space processing according to the frame-level speech text data to obtain semantic embedding space data; Step S24: semantic space mapping is performed on the framed speech signal data by using the semantic embedding space data to generate framed speech semantic coordinate data; Step S25: performing time-series semantic trajectory processing according to the framed speech semantic coordinate data to obtain speech semantic trajectory data.

4. The method for reducing noise of a speech signal according to claim 3, characterized in that: Step S23 includes the following steps: Step S231: converting the frame-level speech text data into a high-dimensional semantic vector to generate semantic original vector data; Step S232: constructing a high-dimensional manifold structure according to the semantic original vector data to obtain semantic manifold structure data; Step S233: using the frame-level speech text data to identify the core semantic factors of the semantic manifold structure data, and generating core semantic factor data; Step S234: using a preset speech corpus to perform contextual semantic embedding matching on the frame-level speech text data, and generating semantic corpus through a generative adversarial network to obtain diversified corpus data; Step S235: decoupling semantic features according to the diversified corpus data, and performing semantic association network processing to generate semantic relationship network data; Step S236: Dynamically semantically fuse the semantic manifold structure data through the semantic relationship network data to obtain dynamic semantic embedding data; Step S237: semantically calibrate the dynamic semantic embedding data, and construct a KD tree fast index based on the core semantic factor data, so as to obtain semantic embedding space data.

5. The method for reducing noise of a speech signal according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: constructing a semantic vector field according to the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; Step S32: performing grid tensor feature analysis on the speech semantic vector field data to generate grid semantic tensor data; performing semantic vector streamline tracing on the speech semantic vector field data through the grid semantic tensor data to obtain semantic vector streamline data; Step S33: performing streamline stability evaluation according to the semantic vector streamline data to generate streamline stability evaluation data; Step S34: performing semantic eddy current detection on the speech semantic vector field data based on the semantic vector streamline data to generate speech eddy current area data; Step S35: Calculating the average speed of the semantic vector streamline data through a preset multi-scale sliding window to generate semantic streamline speed sequence data; Step S36: extracting semantic acceleration from speech eddy current region data based on semantic streamline velocity sequence data, and quantifying eddy current intensity to generate eddy current intensity data; Step S37: Mark the speech noise area of ​​the speech semantic vector field data using the eddy current intensity data, the speech eddy current area data, and the streamline stability evaluation data to generate semantic noise area data.

6. The method for reducing noise of a speech signal according to claim 5, characterized in that: Step S31 includes the following steps: Step S311: performing semantic space gridding processing on the semantic embedding space data to generate grid semantic space data; performing initial semantic potential energy value calculation on the grid semantic space data to generate initial semantic potential energy data; Step S312: constructing an initial vector field for the grid semantic space data using the initial semantic potential energy data to obtain initial semantic vector field data; Step S313: dynamically adjusting the initial semantic vector field data using the speech semantic trajectory data to generate dynamic semantic vector field data; wherein step S313 is specifically as follows: Step S3131: using the speech semantic trajectory data to perform semantic trajectory mapping processing on the grid semantic space data to generate semantic trajectory grid sequence data; Step S3132: using the grid semantic space data to perform grid elimination on the semantic trajectory grid sequence data to obtain non-semantic trajectory grid data; Step S3133: adjusting the semantic trajectory vector of the initial semantic vector field data using the speech semantic trajectory data based on the semantic trajectory grid sequence data to generate semantic trajectory vector adjustment data; Step S3134: Based on the non-semantic trajectory grid data, the semantic trajectory vector adjustment data is used to perform potential field diffusion simulation on the initial semantic vector field data, and vector adjustment value derivation is performed to generate non-trajectory vector adjustment data; Step S3135: dynamically adjusting the vector field of the initial semantic vector field data by using the semantic trajectory vector adjustment data and the non-trajectory vector adjustment data to generate dynamic semantic vector field data; Step S314: performing vector field smoothing optimization according to the dynamic semantic vector field data to generate speech semantic vector field data.

7. The method for reducing noise of a speech signal according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing noise suppression processing on the speech semantic vector field data through the semantic noise region data, and performing vector field adaptive denoising processing to generate noise-reduced semantic vector field data; Step S42: performing semantic trajectory remapping processing on the noise reduction semantic vector field data using the speech semantic trajectory data to generate semantic trajectory remapping data; Step S43: Dynamic potential energy constraints are performed on the speech semantic vector field data through the semantic trajectory remapping data, and trajectory drift potential energy is calculated to obtain semantic trajectory drift potential energy data; Step S44: Based on a preset noise potential energy threshold, the speech semantic trajectory data is subjected to key semantic noise frame recognition through the semantic trajectory drift potential energy data, and acoustic feature inverse mapping is performed to generate speech key noise frame data; Step S45: Perform noise frame repair processing on the framed speech signal data using the speech key noise frame data to generate noise-reduced speech signal data.

8. The method for reducing noise of a speech signal according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: performing time-frequency analysis on the noise reduction speech signal data to obtain noise reduction speech spectrum data; performing frequency domain conversion processing on the speech semantic vector field data to generate semantic spectrum field data; Step S52: performing multi-base spectrum decomposition processing according to the noise reduction speech spectrum data to generate multi-base frequency speech spectrum data; Step S53: using the semantic spectrum field data to perform semantic harmonic association analysis on the multi-fundamental frequency speech spectrum data to generate semantic harmonic association matrix data; Step S54: performing harmonic semantic confidence evaluation according to the semantic harmonic association matrix data, and performing speech enhancement mask processing to generate speech enhancement guidance mask data; Step S55: semantically enhancing the noise reduction speech spectrum data through the speech enhancement guide mask data, and performing auditory masking spectrum correction to generate corrected speech spectrum data; Step S56: Perform speech signal reconstruction processing according to the modified speech spectrum data to generate enhanced speech signal data.

9. A speech signal noise reduction system, characterized in that: Used to perform the speech signal denoising method according to claim 1, the speech signal denoising system comprises: The adaptive framing module is used to perform adaptive framing processing on the voice signal data to be processed to obtain framed voice signal data; perform fundamental frequency voiceprint feature fusion according to the framed voice signal data to generate initial voiceprint base data; The speech semantic analysis module is used to perform acoustic context enhancement processing based on the initial voiceprint base data to generate acoustic feature vector data; perform semantic embedding space processing on the framed speech signal data through the acoustic feature vector data to obtain semantic embedding space data; perform time-series semantic trajectory processing based on the semantic embedding space data to obtain speech semantic trajectory data; The speech noise recognition module is used to construct a semantic vector field based on the speech semantic trajectory data and the semantic embedding space data to generate speech semantic vector field data; mark the speech noise area of ​​the speech semantic vector field data to generate semantic noise area data; The speech noise repair module is used to perform noise suppression processing on the speech semantic vector field data through the semantic noise area data to generate noise reduction semantic vector field data; perform key semantic noise frame recognition on the speech semantic trajectory data using the noise reduction semantic vector field data to generate speech key noise frame data; perform noise frame repair processing according to the speech key noise frame data to generate noise reduction speech signal data; The speech signal enhancement module is used to perform multi-base spectrum decomposition processing on the noise reduction speech signal data to generate multi-base frequency speech spectrum data; perform auditory masking spectrum correction based on the multi-base frequency speech spectrum data, and perform speech signal reconstruction processing to generate enhanced speech signal data.

10. A speech signal noise reduction device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the speech signal denoising method according to any one of claims 1 to 8 when executing the computer program.

Images