A speech noise reduction system in a multi-noise environment

Abstract

The invention discloses a speech noise reduction system in a multi-noise environment, and specifically relates to the technical field of speech noise reduction, comprising a basic division module that divides speech data into a plurality of basic speech segments of the same duration; a rationality analysis module that determines whether the segmentation points of adjacent segments are reasonable, and if not, an optimization marking module that finds new segmentation points for optimization; a loop module that repeatedly verifies the optimization result until all speech segments are segmented reasonably; a noise reduction module that extracts speech and noise features, and uses a noise reduction strategy to select a model for noise reduction processing; a reorganization module that reorganizes the noise-reduced speech segments into complete speech outputs; the invention divides a speech signal into a plurality of segments, and can match the most suitable noise reduction algorithm for each segment according to its noise features, and can cope with a more complex and unstable noise environment through a dynamically adapted noise reduction strategy, and ensure that the speech quality is well improved in various scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of speech noise reduction, and more particularly to a speech noise reduction system in a multi-noise environment. Background Art

[0002] Speech noise reduction technology is mainly used to improve the quality of speech signals in noisy environments. With the widespread application of communication technology and speech recognition technology, people have higher and higher requirements for the quality of speech signals. However, in practical applications, speech signals are usually interfered by various noise sources, such as environmental noise, background music, traffic noise, industrial noise, etc. These noises will significantly affect the clarity and intelligibility of speech signals, bringing great challenges to speech communication, speech recognition and speech enhancement.

[0003] At present, traditional speech noise reduction technology mainly processes a single noise source, but in actual environments, speech signals are often interfered by multiple noises at the same time, and the spectral characteristics of these noises may overlap or be similar, resulting in unsatisfactory noise reduction effects of existing technologies in multi-noise environments. Especially in complex multi-noise environments, a single noise reduction method often cannot effectively suppress the interference of different types of noise on speech signals.

[0004] In order to solve this problem, in recent years, more and more research has focused on speech noise reduction technology in multi-noise environments. Through methods such as multi-channel signal processing, deep learning, and blind source separation, speech signals and various noise signals can be more accurately identified and separated, thereby achieving better noise reduction effects. However, these methods still face problems such as high processing complexity and poor adaptability to noise processing. Therefore, the present invention proposes a speech noise reduction system in a multi-noise environment, in order to solve the above problems by dividing reasonable speech segments and matching accurate noise reduction strategies for them. Summary of the invention

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A speech noise reduction system in a multi-noise environment, comprising a basic division module, a rationality analysis module, an optimization marking module, a circulation module, a noise reduction module, and a reorganization module;

[0007] The basic segmentation module is used to obtain input voice data and segment the voice data according to basic rules to generate multiple voice segments of the same duration. Each voice segment is used as a basic voice segment and as a basic unit for subsequent processing.

[0008] The rationality analysis module determines whether the time point of segmentation is reasonable by analyzing adjacent basic speech segments;

[0009] The optimization marking module is used for finding an optimization marking point as a new segmentation point between two unreasonable basic voice segments when the rationality analysis module detects that the segmentation time of adjacent basic voice segments is unreasonable, and performing a segmentation operation to obtain two optimized voice segments. If no optimization marking point is found, the two adjacent basic voice segments are obtained as one optimized voice segment.

[0010] The loop module is used to loop-verify the rationality of the latest divided optimized speech segment results through the rationality analysis module and the optimization marking module until the latest divided optimized speech segment results are all reasonable, and all the noise-reduced speech segments after loop confirmation are obtained;

[0011] The noise reduction module is used to extract noise features and speech quality features for each noise reduction speech segment, obtain a noise feature group and a speech quality feature group respectively, and then substitute them into the pre-trained noise reduction strategy selection model, select the corresponding noise reduction strategy according to the output result of the noise reduction strategy selection model to perform noise reduction processing, and obtain all the speech processing segments after noise reduction processing;

[0012] The reorganization module is used to reorganize all speech processing segments to obtain complete speech output.

[0013] In a preferred embodiment, the rationality analysis module analyzes adjacent basic speech segments, which includes:

[0014] Analyze amplitude changes and determine amplitude mutations before and after the cutting points of adjacent fragments;

[0015] Analyze energy changes and determine the energy difference before and after the cutting point;

[0016] Analyze the change of signal-to-noise ratio and calculate the change of signal-to-noise ratio before and after the cutting point;

[0017] and inspection and analysis operations.

[0018] In a preferred embodiment, when the rationality analysis module is used, the absolute value of the difference in amplitude mutation before and after the cutting point of adjacent fragments, the absolute value of the difference in energy before and after the cutting point, and the absolute value of the difference in signal-to-noise ratio before and after the cutting point are obtained;

[0019] When the absolute value of the difference in amplitude mutations between adjacent fragment cutting points is greater than or equal to the preset reference threshold 1, an abnormal signal is generated; when the absolute value of the energy difference before and after the cutting point is greater than or equal to the preset reference threshold 2, an abnormal signal is generated; when the absolute value of the difference in signal-to-noise ratio before and after the cutting point is greater than or equal to the preset reference threshold 3, an abnormal signal is generated;

[0020] The test analysis operation refers to judging whether the difference in the mean noise intensity of two windows is significant through the t test or judging whether the frequency distribution of each noise intensity interval is significantly different through the chi-square test. If the difference in the mean noise intensity is significant or the frequency distribution is significantly different, an abnormal signal is generated;

[0021] In a preferred embodiment, when an abnormal signal is generated, it indicates that the time point of segmentation is reasonable.

[0022] When looking for an optimized marking point as a new segmentation point between two unreasonable basic speech segments, the amplitude change, energy change and signal-to-noise ratio change before and after each segmentation time point are explored, and the intensity of these three changes is summarized to generate a change intensity index. Then, the segmentation time point corresponding to the maximum value of the change intensity index is used as an alternative marking point, and the change intensity index of the alternative marking point is compared with the preset change intensity threshold. If the change intensity index of the alternative marking point is greater than the preset change intensity threshold, the alternative marking point is converted into an optimized marking point. If the change intensity index of the alternative marking point is less than or equal to the preset change intensity threshold, the alternative marking point is discarded, that is, the optimized marking point cannot be found.

[0023] In a preferred embodiment, the formula for generating the severe change index is:

[0024] ;

[0025] , , are all preset non-zero scale factors, , , are all preset non-zero adjustment coefficients for regulating the impact of each difference term. is a preset non-zero constant, is the change intensity index, is the absolute value of the average amplitude difference of the segments before and after the cutting point, is the absolute value of the energy difference between the fragments before and after the cutting point, It is the absolute value of the difference in signal-to-noise ratio between the fragments before and after the cutting point.

[0026] In a preferred embodiment, the noise feature group includes noise spectrum features, noise energy features, and noise waveform features.

[0027] In a preferred embodiment, the speech quality feature group includes a speech length feature and a speech quality index feature.

[0028] In a preferred embodiment, the logic for obtaining the voice quality index is:

[0029] The average value of the short-time signal-to-noise ratio, the average value of the instantaneous clarity, and the average value of the spectrum entropy of the noise-reduced speech segment are obtained, and then weighted summation is performed to obtain the speech quality index corresponding to the noise-reduced speech segment.

[0030] In a preferred embodiment, the denoising strategy selection model is a convolutional neural network model.

[0031] Technical effects and advantages of the present invention:

[0032] Traditional speech noise reduction methods usually process the entire speech signal as a whole. This processing method easily leads to unstable noise reduction effects or excessive noise reduction (i.e., speech signal distortion) in scenarios where the noise type varies greatly. The present invention divides the overall speech signal into multiple smaller speech segments, so that each segment can adopt a personalized noise reduction strategy based on its unique noise characteristics. Specifically, some speech segments may contain strong background noise or have more complex noise sources, while other segments may be relatively clean or have less noise. By performing targeted processing on these segments, the interference of noise on the speech signal can be minimized, and the clarity and intelligibility of the speech signal can be improved.

[0033] Speech signals in a multi-noise environment are often interfered by a variety of noise sources, which may have different intensities and characteristics in different time periods or different frequency bands. By dividing the speech signal into multiple segments, the present invention can match the most suitable noise reduction algorithm for each segment according to the noise characteristics of each segment. The matching of noise reduction strategies for different segments can significantly improve the clarity and intelligibility of the overall speech. The present invention analyzes the changing characteristics of speech and reasonably divides the speech segments so that the boundaries of each segment are in line with the changes in the actual complex environment. A reasonable division strategy can ensure that the subsequent noise reduction algorithm can accurately focus on the target noise source when processing each segment to avoid accidentally damaging the speech signal itself. In practical applications, this precise segment division will significantly reduce the generation of "pseudo-noise", making the denoised speech more natural and smooth, while also avoiding speech distortion or blurring caused by excessive noise reduction.

[0034] In actual noise environments, noise usually changes dynamically. Traditional noise reduction methods often use fixed parameters and strategies, which may not be able to adapt to rapidly changing noise environments. However, the present invention can evaluate the noise characteristics of each segment in real time and automatically adjust the noise reduction strategy by dividing and dynamically analyzing the speech segments. Through the dynamically adaptive noise reduction strategy, the present invention can cope with more complex and unstable noise environments, ensuring that the voice quality is well improved in various scenarios.

[0035] After dividing the speech signal into multiple segments, the system can apply different noise reduction algorithms to each segment, which not only improves the noise reduction accuracy, but also avoids excessive calculations on the entire speech signal. Through targeted segment processing, unnecessary calculations can be effectively reduced, reducing the computational complexity of the system, so that it can run efficiently on devices with limited resources (such as mobile devices and embedded systems). BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0037] Figure 1 The schematic diagram of a speech noise reduction system in a multi-noise environment in the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Reference Figure 1 The following embodiments are obtained:

[0040] Example 1

[0041] In a speech processing system, the design idea of ​​dividing speech into multiple small speech segments and gradually processing them is usually to improve processing efficiency and processing accuracy, while making the operation of each stage more targeted, but how to ensure the rationality of the division is a problem that needs to be solved. Therefore, the present invention proposes a speech noise reduction system in a multi-noise environment, including a basic division module, a rationality analysis module, an optimization marking module, a circulation module, a noise reduction module, and a reorganization module; the basic division module, the rationality analysis module, the optimization marking module, the circulation module, the noise reduction module, and the reorganization module are communicated and connected.

[0042] The basic division module is used to obtain the input speech data and divide the speech data according to the basic rules to generate multiple speech segments of the same length. Each speech segment is used as a basic speech segment and as the basic unit for subsequent processing. The main function of this module is to divide the input continuous speech data into several small speech segments of the same length according to the set basic rules. Each speech segment is regarded as a basic unit for processing in the subsequent processing. The purpose of basic division is to divide the original long speech signal into smaller units that are easier to process, so as to avoid wasting computing resources or processing difficulties due to too long speech during the processing process. The purpose of dividing the speech segments is:

[0043] Segmentation improves processing efficiency: Large-scale speech data is difficult to process efficiently in one step. By dividing it into multiple segments, the amount of data for each calculation can be reduced, thereby increasing processing speed.

[0044] Local feature extraction: Each speech segment can extract speech features independently, which facilitates detailed analysis and optimization of different features.

[0045] Facilitates noise reduction processing: After dividing the voice data into small segments, noise reduction processing can be performed on each segment separately to reduce the impact of noise accumulation over a long period of time.

[0046] The rationality analysis module analyzes adjacent basic speech segments to determine whether the time point of segmentation is reasonable. This module is responsible for analyzing adjacent basic speech segments to check whether the time point of segmentation is reasonable. The purpose of rationality analysis is to improve the processing quality. Unreasonable segmentation may lead to errors in subsequent feature extraction and processing, affecting the overall effect.

[0047] The optimization marking module is used to find the optimization marking point as a new segmentation point between the two unreasonable basic voice segments when the rationality analysis module detects that the segmentation time of adjacent basic voice segments is unreasonable, and perform segmentation operations to obtain two optimized voice segments. If the optimization marking point cannot be found, the two adjacent basic voice segments are obtained as one optimized voice segment; when the rationality analysis module detects that the segmentation point is unreasonable, the module will re-segment the voice segment by finding a new optimization marking point. If the optimization point cannot be found, the adjacent segments will be merged to ensure the rationality of each processing unit. The optimization marking module is to correct unreasonable segmentation. Sometimes, the basic division may occur in an inappropriate place. At this time, the optimization marking module can correct the error and find a more suitable segmentation point, so as to find the most matching noise reduction strategy for each segment, and different noise reduction strategies perform noise reduction processing on the corresponding segments respectively, which can reduce the overall noise reduction time and have better noise reduction accuracy.

[0048] The loop module is used to loop-verify the rationality of the latest divided optimized speech segments through the rationality analysis module and the optimization marking module until all the latest divided optimized speech segments are reasonable and all the noise-reduced speech segments are confirmed by the loop; the loop module repeatedly verifies whether the optimized speech segments are reasonable until all the optimized speech segments meet the rationality requirements. This module ensures that the final speech division is both reasonable and accurate. The necessity of loop verification: Improve system stability: Through repeated verification, the system can handle edge cases or complex scenarios to ensure that the speech segmentation is more in line with the speech processing rules. Correct accumulated errors: In multiple rounds of loops, the system can gradually correct errors to ensure that there is no obvious unreasonable segmentation in the final output.

[0049] The noise reduction module is used to extract noise features and voice quality features for each noise reduction speech segment, obtain noise feature groups and voice quality feature groups respectively, and then substitute them into the pre-trained noise reduction strategy selection model, select the corresponding noise reduction strategy for noise reduction processing according to the output result of the noise reduction strategy selection model, and obtain all the speech processing segments after noise reduction processing; this module extracts noise features and voice quality features for each noise reduction speech segment, and inputs these features into the pre-trained noise reduction strategy selection model to select a suitable noise reduction strategy, and then performs noise reduction processing on the speech segment according to the selected strategy. Improve speech quality. Noise in speech may affect the final effect. The noise reduction module can effectively remove background noise and improve the clarity of speech. Through targeted processing, different noise types and speech quality require different noise reduction strategies. The noise reduction module achieves accurate noise reduction through feature extraction and strategy selection.

[0050] The reorganization module is used to reorganize all speech processing segments to obtain complete speech output. The speech segments after noise reduction processing will be reorganized into complete speech output. The reorganization module merges all processed speech segments into a continuous speech stream without obvious breaks. The purpose of the reorganization module is to restore integrity. After each speech segment is processed, it must be reassembled to restore it to a complete and coherent speech file to ensure the naturalness and fluency of the system output. To maintain the fluency of the original speech, the reorganization process needs to ensure that the speech segments are connected naturally to avoid sounding like a spliced ​​and patchwork product.

[0051] The rationality analysis module analyzes adjacent basic speech segments by: analyzing amplitude changes to determine the amplitude mutation before and after the cutting point of adjacent segments; analyzing energy changes to determine the energy difference before and after the cutting point; analyzing signal-to-noise ratio changes to calculate the changes in signal-to-noise ratio before and after the cutting point; and verifying the analysis operation.

[0052] When the rationality analysis module is used, the absolute value of the difference in amplitude mutation before and after the cutting point of adjacent fragments, the absolute value of the difference in energy before and after the cutting point, and the absolute value of the difference in signal-to-noise ratio before and after the cutting point are obtained;

[0053] When the absolute value of the difference in amplitude mutations between adjacent fragment cutting points is greater than or equal to the preset reference threshold 1, an abnormal signal is generated; when the absolute value of the energy difference before and after the cutting point is greater than or equal to the preset reference threshold 2, an abnormal signal is generated; when the absolute value of the difference in signal-to-noise ratio before and after the cutting point is greater than or equal to the preset reference threshold 3, an abnormal signal is generated;

[0054] The test analysis operation refers to judging whether the difference in the mean noise intensity of two windows is significant through the t test or judging whether the frequency distribution of each noise intensity interval is significantly different through the chi-square test. If the difference in the mean noise intensity is significant or the frequency distribution is significantly different, an abnormal signal is generated;

[0055] When an abnormal signal is generated, it means that the time point of segmentation is reasonable.

[0056] The core function of the rationality analysis module is to conduct a detailed analysis of adjacent basic speech segments to verify the rationality of the speech cutting points and ensure the accuracy of the cutting process. The cutting points are usually determined when sudden noise or abnormal changes in the speech signal occur, and the change in the noise environment is judged to start or end. By analyzing the signal characteristics before and after these cutting points, abnormal changes in the speech signal can be effectively identified and marked, thereby helping to optimize the speech processing process, especially in a variable noise environment.

[0057] Amplitude change analysis: Amplitude mutation refers to abnormal changes in signal amplitude before and after the cutting point. The cutting points of adjacent speech segments are usually disturbed by sudden noise, which may cause the amplitude to change significantly in an instant. In the rationality analysis module, by calculating the amplitude difference of adjacent speech segments, it can be determined whether the amplitude change before and after the cutting point is within the normal fluctuation range. If the absolute value of the amplitude mutation exceeds the preset baseline threshold of one, it will be marked as an abnormal signal. This method can effectively capture the drastic amplitude changes caused by sudden noise and help the system identify the corresponding cutting point when the noise environment changes.

[0058] Energy change analysis: Energy change analysis focuses on whether there are abnormal fluctuations in the energy level of the speech signal. In speech signal processing, energy is an important indicator to measure speech strength. The energy mutation before and after the cutting point may be caused by noise or environmental changes. By calculating the energy difference before and after the cutting point and comparing it with the set reference threshold 2, if the difference exceeds the threshold, it will trigger the generation of abnormal signals. Energy difference analysis can effectively identify changes in signal strength caused by noise or interference, ensuring the rationality of the cutting point.

[0059] Analysis of signal-to-noise ratio changes: The signal-to-noise ratio is an important indicator for measuring the quality of speech signals, representing the ratio between speech signals and background noise. The rationality analysis module evaluates whether the cutting point is in a reasonable noise environment by calculating the difference in the signal-to-noise ratio before and after the cutting point. If the absolute value of the change in the signal-to-noise ratio before and after the cutting point exceeds the preset benchmark threshold of three, it is considered that there is an abnormal signal at the cutting point. Through this analysis, the signal-to-noise ratio anomaly caused by the drastic change in noise intensity can be detected, which helps to determine whether the cutting point is affected by interference or noise, and the starting or ending time of the change in the noise environment corresponding to the cutting point.

[0060] Test and analysis operation: Test and analysis operation is the statistical test link in the rationality analysis module, which aims to verify whether there is a significant difference between the signals before and after the cutting point through more scientific statistical methods. Commonly used test methods include t test and chi-square test:

[0061] t-test: used to determine whether the difference in the mean noise intensity of the two windows before and after the cut point is significant. If the difference in the mean noise intensity of the two windows exceeds the normal fluctuation range, the t-test will show significance, thus determining that the cut point may be caused by abnormal noise.

[0062] Chi-square test: used to test whether the frequency distribution of different noise intensity intervals is significantly different. Through the chi-square test, the system can determine whether there is a significant change in the frequency distribution of different noise intensity intervals before and after the cut point. If there is a significant difference, it means that the distribution of noise intensity is abnormal, and then the rationality of the cut point is determined.

[0063] If these statistical tests indicate that the noise intensity before and after the cut point is significantly different, the system will generate an abnormal signal, which may be caused by sudden noise or environmental changes, indicating the rationality of the cut point.

[0064] Abnormal signal generation and judgment of rationality of cutting point: Through the above analysis, when the amplitude change, energy change, signal-to-noise ratio change and other indicators before and after the cutting point are abnormal, or the statistical test results show significant differences, the rationality analysis module will generate abnormal signals. These abnormal signals indicate that the current cutting point does not conform to the change law of normal speech signals, which may be caused by noise, interference or sudden environmental changes. Therefore, when an abnormal signal appears, the system can judge that the cutting point is reasonable, ensuring that the speech segment obtained by segmentation at the position of the cutting point can better meet the expected effect of speech processing in the subsequent noise reduction processing.

[0065] When looking for an optimized marking point as a new segmentation point between two unreasonable basic speech segments, the amplitude change, energy change and signal-to-noise ratio change before and after each segmentation time point are explored, and the intensity of these three changes is summarized to generate a change intensity index. Then, the segmentation time point corresponding to the maximum value of the change intensity index is used as an alternative marking point, and the change intensity index of the alternative marking point is compared with the preset change intensity threshold. If the change intensity index of the alternative marking point is greater than the preset change intensity threshold, the alternative marking point is converted into an optimized marking point. If the change intensity index of the alternative marking point is less than or equal to the preset change intensity threshold, the alternative marking point is discarded, that is, the optimized marking point cannot be found.

[0066] The formula for generating the severity index is:

[0067] ;

[0068] , , are all preset non-zero scale factors, , , are all preset non-zero adjustment coefficients for regulating the impact of each difference term. is a preset non-zero constant, is the change intensity index, is the absolute value of the average amplitude difference of the segments before and after the cutting point, is the absolute value of the energy difference between the fragments before and after the cutting point, It is the absolute value of the difference in signal-to-noise ratio between the fragments before and after the cutting point.

[0069] In the process of audio signal processing, the segmentation points of the original audio clips are often not the most appropriate. These segmentation points may be generated based on simple time intervals or voice activity detection criteria, but may not conform to the characteristics of the speech signal of the audio content. Therefore, it is particularly important to find a more accurate optimization mark point.

[0070] By introducing the changes in amplitude, energy, and signal-to-noise ratio as measurement criteria, we can comprehensively consider different aspects of the speech signal and select the position that best represents the "natural split point". These three characteristics of the audio signal (amplitude change, energy change, and signal-to-noise ratio change) directly reflect the dynamic changes of the signal itself, and are critical for judging the rationality of the speech segment split point.

[0071] Amplitude change: Amplitude (or amplitude) is a reflection of the strength of an audio signal. When an audio signal undergoes a sudden amplitude change, it often means that there is an important change in the audio signal. By observing the change in amplitude, you can find significant turning points in the audio signal.

[0072] Energy change: The energy of an audio signal is another manifestation of signal strength. Energy fluctuations often reflect changes in the rhythm of a speech signal or the strength of a speech. Strong energy changes are often associated with important events in speech, helping to identify appropriate segmentation points.

[0073] Signal-to-noise ratio change: Signal-to-noise ratio is an important indicator to measure the clarity of speech signals. Changes in signal-to-noise ratio can reflect changes in background noise or interference. A dramatic change in signal-to-noise ratio may mean that the speech signal has shifted from one state to another, which is a potential cutoff point.

[0074] By analyzing the degree of change in amplitude, energy, and signal-to-noise ratio before and after each segmentation time point, a "drastic change index" can be calculated. This index is actually a comprehensive indicator that quantifies the changes in these features, and can reveal which signal changes are the largest before and after the segmentation points. Generally speaking, drastic changes mean that the structure of the signal has changed significantly, which is often the manifestation of a natural segmentation point. A large drastic change index often means that the point is at an important turning point in the speech signal, while a smaller change may be background noise or a weak change, which is not suitable as an optimized segmentation point.

[0075] By comparing the "drastic change index" of the candidate segmentation points with the preset drastic change threshold, the segmentation points that truly meet the optimization requirements can be effectively screened out. If the drastic change index of the candidate marking point is greater than the preset threshold, it means that the segmentation point is at the turning point of the signal, meets the optimization criteria, and can be used as an optimization marking point; on the contrary, if its drastic change index is less than or equal to the threshold, it means that the change before and after the segmentation point is not enough to consider it as a natural and meaningful segmentation position, and then the point is discarded.

[0076] Improve segmentation accuracy: By selecting markers with large change indices, we can avoid loss or distortion of speech information due to incorrect segmentation or over-segmentation. Accurate segmentation not only improves the accuracy of speech recognition, but also makes subsequent speech processing (such as speech synthesis) more natural and efficient.

[0077] Improve speech quality: Reasonable segmentation points help to restore or retain speech features in the process of multiple noise changes, making the recognized speech after noise reduction closer to natural human speech and reducing the unnatural feeling in recognition.

[0078] Adapt to noisy environment: In noisy environments or environments with large background noise, the judgment of signal-to-noise ratio changes is particularly critical. By optimizing the selection of marker points based on the changes in the signal-to-noise ratio, the interference of noise on speech segmentation can be effectively reduced, and the segmentation effect in complex environments can be improved.

[0079] Automation and Optimization: This method not only reduces manual intervention, but also automatically optimizes the audio segmentation process by setting reasonable thresholds. It has important practical significance for large-scale audio processing tasks (such as automatic transcription, speech recognition, etc.).

[0080] The noise feature group includes noise spectrum features, noise energy features, and noise waveform features. Noise spectrum features refer to the spectrum information of noise obtained through frequency domain analysis. It reflects the energy distribution of noise at different frequencies. Spectral features help distinguish the spectrum differences between speech signals and noise. Noise usually has different frequency components and spectral distributions. Understanding these differences can help algorithms identify and suppress noise. For example, in noise suppression and speech enhancement, noise spectrum features can help distinguish noise from speech signals, thereby achieving more accurate noise removal.

[0081] The noise energy feature refers to the energy of the noise part of the signal, which is usually calculated by analyzing the signal in the time domain or frequency domain. The energy feature helps to quantify the intensity of the noise. Strong noise may seriously affect the voice quality, and understanding its energy helps to effectively suppress it. This feature is usually used in speech enhancement and noise suppression algorithms to dynamically adjust the denoising strategy and adjust the filter parameters according to the noise intensity to achieve the best effect.

[0082] Noise waveform features refer to the time domain waveform properties of noise, such as the degree of fluctuation and periodicity of the waveform. Waveform features help describe the temporal structure of noise. Different types of noise (such as white noise, factory noise, wind noise, etc.) have different time domain waveforms. Waveform features can help identify the type of noise. This feature plays an important role in speech enhancement and noise classification. By analyzing the noise waveform, the system can adapt to the noise characteristics in different environments and further optimize the denoising process.

[0083] The speech quality feature group includes speech length feature and speech quality index feature. The speech length feature refers to the duration of the speech signal, usually in seconds. The length of speech affects the task of speech recognition. Shorter speech may lack sufficient contextual information, and longer speech may contain more noise or redundant information. Understanding the speech length helps adjust the system's processing strategy. In tasks such as speech recognition and speech synthesis, the speech length feature can help the system optimize the response speed and accuracy of the model when processing speech of different lengths.

[0084] The logic for obtaining the voice quality index is as follows: obtain the short-term signal-to-noise ratio average, instantaneous clarity average, and spectral entropy average of the noise-reduced voice segment, and then perform weighted summation to obtain the voice quality index corresponding to the noise-reduced voice segment. The signal-to-noise ratio reflects the ratio between the voice signal and the noise. The higher the signal-to-noise ratio, the clearer the voice signal and the smaller the noise impact. Therefore, a higher signal-to-noise ratio helps improve the voice quality index. Clarity indicates the clarity of the voice, which is usually related to the intelligibility and distortion of the voice. Higher clarity means that the voice is easier to understand and there is less noise interference. Spectral entropy measures the complexity or amount of information in the voice signal. A lower spectral entropy value usually means that the spectral structure of the voice signal is relatively simple and regular, while a higher spectral entropy may indicate that the signal has more noise or distortion and the voice quality is poor.

[0085] A higher SQI usually means better noise reduction, clearer voice signals and less noise interference, and higher overall voice quality, suitable for listening and understanding. A lower SQI means poorer noise reduction, and the voice signal may be masked by noise or severely distorted, resulting in lower voice quality and may be difficult to understand clearly.

[0086] The noise reduction strategy selection model is a convolutional neural network model. In speech signal processing tasks, the goal of the noise reduction strategy selection model is to select a suitable noise reduction strategy based on the input noise characteristics and speech quality characteristics, thereby optimizing the noise reduction effect and improving the clarity and quality of speech. The convolutional neural network (CNN) is used as the noise reduction strategy selection model because CNN can efficiently extract local features and can handle complex noise patterns. The significance of noise reduction strategy selection:

[0087] Improve speech quality: The noise reduction effect varies greatly in different noise environments. The appropriate noise reduction strategy can significantly improve the clarity and intelligibility of speech.

[0088] Reduce computational complexity: Select appropriate noise reduction strategies based on noise characteristics and speech quality characteristics, avoiding excessive computation and resource waste caused by using global noise reduction strategies.

[0089] Adapt to various noise environments: There are many kinds of noise environments, and through intelligent selection strategies, it can be optimized for different noise types.

[0090] Noise reduction strategy selection process: Input feature extraction: First, two feature groups need to be extracted: a noise feature group and a speech quality feature group.

[0091] Convolutional neural network processing: The noise feature group and the speech quality feature group are input into the pre-trained convolutional neural network model. Through multi-layer convolution operations, CNN can automatically learn important patterns from the original features and output a decision for noise reduction strategy selection.

[0092] Select a noise reduction strategy: Based on the output of the CNN model, select the noise reduction strategy that best suits the current noise environment and speech quality. For example, if the noise characteristics indicate low-frequency noise and poor speech quality, a specific filtering strategy may be selected.

[0093] Apply the selected strategy: According to the strategy output by the model, select the corresponding noise reduction method and apply it to the input speech signal, and finally obtain a processed clear speech signal.

[0094] Examples of noise reduction strategies: Spectral subtraction: Applicable to scenarios with relatively stable background noise, especially white noise. It reduces noise interference by estimating the noise spectrum and subtracting it from the speech signal.

[0095] Wiener filter: A dynamically adjusted filter based on the signal-to-noise ratio, suitable for scenarios with low signal-to-noise ratios. It adjusts the filter gain based on the strength of the signal and the strength of the noise to preserve the clarity of the speech signal as much as possible.

[0096] Deep learning denoising: For complex noise (such as noise in a dynamic environment), deep learning models can learn denoising strategies for various noises through a large amount of training data. CNN can automatically learn noise patterns and generate the best filter in this process.

[0097] Blind source separation: This strategy can separate multiple sound sources and is widely used in multi-microphone arrays. It can recover the signals of each source from mixed speech and noise.

[0098] The goal of the noise reduction strategy selection model is to automatically select the most appropriate noise reduction method based on noise characteristics and speech quality characteristics, so as to obtain the best speech quality in different noise environments. Convolutional neural networks play a key role in automatic feature learning and decision support in this process. By intelligently selecting noise reduction strategies, the clarity of speech signals can be maximized, improving the performance of applications such as speech recognition.

[0099] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0100] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0101] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0102] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0103] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A speech noise reduction system in a multi-noise environment, characterized in that: It includes basic division module, rationality analysis module, optimization marking module, circulation module, noise reduction module and reorganization module; The basic segmentation module is used to obtain input voice data and segment the voice data according to basic rules to generate multiple voice segments of the same duration. Each voice segment is used as a basic voice segment and as a basic unit for subsequent processing. The rationality analysis module determines whether the time point of segmentation is reasonable by analyzing adjacent basic speech segments; The optimization marking module is used for finding an optimization marking point as a new segmentation point between two unreasonable basic voice segments when the rationality analysis module detects that the segmentation time of adjacent basic voice segments is unreasonable, and performing a segmentation operation to obtain two optimized voice segments. If no optimization marking point is found, the two adjacent basic voice segments are obtained as one optimized voice segment. The loop module is used to loop-verify the rationality of the latest divided optimized speech segment results through the rationality analysis module and the optimization marking module until the latest divided optimized speech segment results are all reasonable, and all the noise-reduced speech segments after loop confirmation are obtained; The noise reduction module is used to extract noise features and speech quality features for each noise reduction speech segment, obtain a noise feature group and a speech quality feature group respectively, and then substitute them into the pre-trained noise reduction strategy selection model, select the corresponding noise reduction strategy according to the output result of the noise reduction strategy selection model to perform noise reduction processing, and obtain all the speech processing segments after noise reduction processing; The reorganization module is used to reorganize all speech processing segments to obtain complete speech output; When judging whether the time point of segmentation is reasonable, a dramatic change index is generated, and the generation formula is: ； are all preset non-zero scale factors, are all preset non-zero adjustment coefficients for regulating the impact of each difference term. is a preset non-zero constant, is the change intensity index, is the absolute value of the average amplitude difference of the segments before and after the cutting point, is the absolute value of the energy difference between the fragments before and after the cutting point, is the absolute value of the difference in signal-to-noise ratio between the fragments before and after the cutting point; The noise feature group includes noise spectrum features, noise energy features, and noise waveform features; The speech quality feature group includes speech length feature and speech quality index feature; The logic for obtaining the voice quality index is: Obtaining the average value of the short-time signal-to-noise ratio, the average value of the instantaneous clarity, and the average value of the spectrum entropy of the noise-reduced speech segment, and then performing weighted summation to obtain the speech quality index corresponding to the noise-reduced speech segment; When looking for an optimized marking point as a new segmentation point between two unreasonable basic speech segments, the amplitude change, energy change and signal-to-noise ratio change before and after each segmentation time point are explored, and the intensity of these three changes is summarized to generate a change intensity index. Then, the segmentation time point corresponding to the maximum value of the change intensity index is used as an alternative marking point, and the change intensity index of the alternative marking point is compared with the preset change intensity threshold. If the change intensity index of the alternative marking point is greater than the preset change intensity threshold, the alternative marking point is converted into an optimized marking point. If the change intensity index of the alternative marking point is less than or equal to the preset change intensity threshold, the alternative marking point is discarded, that is, the optimized marking point cannot be found.

2. The speech noise reduction system in a multi-noise environment according to claim 1, characterized in that: The rationality analysis module analyzes adjacent basic speech segments: Analyze amplitude changes and determine amplitude mutations before and after the cutting points of adjacent fragments; Analyze energy changes and determine the energy difference before and after the cutting point; Analyze the change of signal-to-noise ratio and calculate the change of signal-to-noise ratio before and after the cutting point; and inspection and analysis operations.

3. The speech noise reduction system in a multi-noise environment according to claim 2, characterized in that: When the rationality analysis module is used, the absolute value of the difference in amplitude mutation before and after the cutting point of adjacent fragments, the absolute value of the difference in energy before and after the cutting point, and the absolute value of the difference in signal-to-noise ratio before and after the cutting point are obtained; When the absolute value of the difference in amplitude mutations between adjacent fragment cutting points is greater than or equal to the preset reference threshold 1, an abnormal signal is generated; when the absolute value of the energy difference before and after the cutting point is greater than or equal to the preset reference threshold 2, an abnormal signal is generated; when the absolute value of the difference in signal-to-noise ratio before and after the cutting point is greater than or equal to the preset reference threshold 3, an abnormal signal is generated; The test analysis operation refers to judging whether the difference in the mean noise intensity of two windows is significant through the t test or judging whether the frequency distribution of each noise intensity interval is significantly different through the chi-square test. If the difference in the mean noise intensity is significant or the frequency distribution is significantly different, an abnormal signal is generated; When an abnormal signal is generated, it means that the time point of segmentation is reasonable.

4. The speech noise reduction system in a multi-noise environment according to claim 3, characterized in that: The denoising strategy selection model is a convolutional neural network model.

Images