An adaptive noise reduction method, system and device

Abstract

The present application relates to the field of noise reduction technology, in particular to a kind of adaptive noise reduction method, system and equipment, wherein the method comprises: obtaining mixed audio signal and dividing into short-time segment, after normalizing short-time segment, the normalized short-time energy and normalized change intensity of short-time segment are calculated, the short-time segment after normalization is converted into spectrogram and input into light time-frequency feature extraction network, and noise state feature set is output by extraction network;The comprehensive score of short-time segment in noise state feature set is calculated, the starting segment of change, the change influence range and the change type are determined, and the state change description of noise is constructed;The normalized processing intensity of state change description is calculated, noise reduction execution item is constructed, noise reduction execution scheme is generated by sorting noise reduction execution item;For each noise reduction execution item in noise reduction execution scheme, time-domain audio is overlapped and spliced according to the original time sequence of short-time segment, and noise reduction output signal is output.

Description

Technical Field

[0001] This invention relates to the field of noise reduction technology, and more particularly to an adaptive noise reduction method, system, and device. Background Technology

[0002] In fire rescue, police duty, public safety response, aviation ground support, and various special operations scenarios, communication headsets, throat vibration headsets, bone conduction headsets, aviation headsets, and walkie-talkie accessories are responsible for the real-time transmission of on-site instructions, status information, and coordination commands. Their noise reduction targets differ significantly from those of ordinary consumer headsets. These devices typically operate in environments with multiple sound sources, including fans, engines, vehicles, breathing apparatus, sirens, mechanical collisions, strong winds, and near-field shouting by personnel. The noise is not only high in intensity but also exhibits significant non-stationarity and abruptness. Common manifestations include the long-term presence of continuous low-frequency noise, a rapid increase in wind noise, broadband impacts from metal collisions, spectral pattern switching caused by equipment startup, and changes in pickup patterns due to changes in personnel posture or wearing position. Existing noise reduction technologies in professional communication equipment typically rely on fixed filter parameters, continuous adjustment of short-term energy or spectral statistics, mode switching under preset noise scenarios, or the use of conventional speech enhancement algorithms to suppress noise in the current audio segment.

[0003] These technologies can achieve certain results in relatively stable background noise, but they are prone to engineering shortcomings in the rapidly changing noise environment of professional communication sites: First, many algorithms mainly rely on the statistical characteristics of the current segment or the segments that have occurred for processing. When faced with sudden wind noise, impact noise, or sudden equipment startup, the noise reduction action often lags behind the change in noise state. Second, continuous high noise and true state switching may both exhibit high intensity in short-term energy. Simply relying on energy thresholds or fixed spectrum rules can easily confuse stable high noise with effective changes. Third, the processing methods for different noise types vary greatly. Wind noise enhancement is more suitable for continuous low-frequency processing, short-term impact is more suitable for short-range rapid suppression, and noise mode switching requires a smooth redistribution of the frequency band processing ratio. Existing unified adjustment methods are difficult to take into account these scenarios simultaneously. Fourth, professional communication emphasizes speech intelligibility and command continuity. Excessive frequency band suppression, frequent mode jumps, or uneven processing transitions can cause consonant edge weakening, loss of the start word of the call, auditory aspiration, and decreased command recognition.

[0004] Therefore, the adaptive noise reduction problem of this type of equipment is not simply about increasing the noise suppression intensity. Rather, it requires further solutions based on existing technologies such as short-time frequency domain processing, adaptive gain control, and frequency band weighting processing. These solutions address how to continuously characterize the noise state, accurately describe effective changes, transform the results of changes into executable segment-level and frequency band-level processing schemes, and ultimately smoothly apply them to the actual audio signal. This will make the noise reduction process more suitable for the engineering characteristics of professional communication scenarios, such as strong noise, diverse sources, rapid changes, and high requirements for voice transmission. Summary of the Invention

[0005] To address the above problems, this invention provides an adaptive noise reduction method, system, and device.

[0006] To achieve the above objectives, the present invention proposes an adaptive noise reduction method, comprising:

[0007] A mixed audio signal is acquired and divided into short time segments. After normalizing the short time segments, the normalized short-time energy of the short time segments is calculated. Based on the normalized short-time energy, the normalized change intensity between adjacent short time segments is calculated. The normalized short time segments are converted into spectrograms. The spectrograms, normalized short-time energy, and normalized change intensity are input into a lightweight time-frequency feature extraction network. The lightweight time-frequency feature extraction network outputs a noise state feature set.

[0008] A comprehensive score is calculated for short segments in the noise state feature set. Based on the comprehensive score, normalized change intensity, and noise state feature vector corresponding to the short segment, the starting segment of the change, the range of influence of the change, and the change type are determined, and a description of the noise state change is constructed.

[0009] According to the preset execution mapping table, the normalized processing intensity of the state change description is calculated, and the frequency band execution vector of the change type is calculated in combination with the normalized processing intensity. The noise reduction execution items are constructed by combining the state change description, the normalized processing intensity and the frequency band execution vector, and the noise reduction execution items are sorted to generate a noise reduction execution scheme.

[0010] For each noise reduction execution item in the noise reduction execution scheme, the corresponding target processing segment is located from the mixed audio signal, the target processing segment is converted into a short-time spectral representation and the original spectral coefficients are obtained, a transition coefficient is generated according to the change type and the transition mode between short-time segments, the actual spectral retention coefficient is calculated based on the transition coefficient, the actual spectral retention coefficient is applied to the original spectral coefficient corresponding to the target processing segment to generate a corrected spectral coefficient, the corrected spectral coefficients corresponding to all noise reduction execution items are summarized and converted into time-domain audio, the time-domain audio is overlapped and spliced ​​according to the original time order of the short-time segments, and the noise-reduced output signal is output.

[0011] In some embodiments, the lightweight time-frequency feature extraction network includes an input sorting layer, a convolutional processing layer, a gated recurrent processing layer, and an output mapping layer. The sorting layer is used to receive the spectrum, normalized short-time energy, and normalized change intensity and arrange them in temporal order according to short-time segments. The gated recurrent processing layer is used to extract local frequency band texture and expand the observation range of the frequency band. The output mapping layer is used to sort the local frequency band texture and output it as a noise state feature set.

[0012] In some embodiments, determining the starting segment, the range of influence, and the type of change based on the comprehensive score, the normalized change intensity, and the noise state feature vector corresponding to the short-time segment specifically includes:

[0013] A preset number of short time segments are obtained, the comprehensive scores of the short time segments are sorted, and the median level is taken as the natural fluctuation level.

[0014] When the comprehensive score of the short segment to be compared is higher than the sum of the natural fluctuation level and the preset rise threshold, and the change trend of the shape change direction and the normalized change intensity of the corresponding noise state feature vector meets the preset consistency judgment rule, the short segment is determined as the change start segment.

[0015] Continue reading the comprehensive score and noise state feature vector of subsequent short segments in chronological order, and classify the short segments with continuous change features into the same range of change influence;

[0016] When the difference between the comprehensive score of the short-term segment and the natural fluctuation level is less than the threshold, and the noise state feature vector is stable, the short-term segment is determined as the endpoint segment.

[0017] The change type is obtained by combining the shape of the noise state feature vector around the change point, the change trend of the normalized short-time energy, and the fluctuation of the normalized change intensity. The noise state feature vector is obtained through the noise state feature set.

[0018] In some embodiments, converting the normalized short-time segment into a spectrum further includes: if a short-time segment cannot form a valid spectrum due to sampling loss or input buffer abnormality, then generating a corresponding placeholder spectrum using the spectrum size configuration of the previous short-time segment, and recording an abnormal input flag on the short-time segment.

[0019] In some embodiments, the normalized short-time energy is calculated and generated using the sampled values ​​of the short-time segment and the number of sample points.

[0020] In some embodiments, the comprehensive score is calculated and generated by normalizing the change intensity, preset weights, preset weighting coefficients, components of the noise state feature vector, and the number of components.

[0021] In some embodiments, calculating the normalized processing intensity of the state change description according to a preset execution mapping table specifically includes:

[0022] The execution mapping table calls the corresponding type correction value, frequency band processing priority, and voice preservation configuration according to the change type. Based on the type correction value, frequency band processing priority vector, and voice preservation vector, the normalized processing intensity is calculated. The frequency band processing priority vector and voice preservation vector are both written according to the normalization ratio.

[0023] If the change type does not match the execution mapping table, the preset general noise configuration is invoked.

[0024] In some embodiments, the frequency band execution vector is generated by calculating the normalized processing intensity, the frequency band processing priority vector, and the speech preservation vector.

[0025] To achieve the above objectives, another aspect of the present invention proposes an adaptive noise reduction system, comprising:

[0026] The feature extraction module is used to acquire mixed audio signals and divide them into short time segments. After normalizing the short time segments, it calculates the normalized short-time energy of the short time segments, calculates the normalized change intensity between adjacent short time segments based on the normalized short-time energy, converts the normalized short time segments into spectrograms, and inputs the spectrograms, normalized short-time energy, and normalized change intensity into a lightweight time-frequency feature extraction network. The lightweight time-frequency feature extraction network outputs a noise state feature set.

[0027] The state change description module is used to calculate a comprehensive score for short segments in the noise state feature set, and to determine the starting segment, the range of influence and the type of change based on the comprehensive score, the normalized change intensity and the noise state feature vector corresponding to the short segment, and to construct a state change description of the noise.

[0028] The noise reduction scheme generation module is used to calculate the normalized processing intensity of the state change description according to a preset execution mapping table, calculate the frequency band execution vector of the change type in combination with the normalized processing intensity, construct noise reduction execution items in combination with the state change description, normalized processing intensity and frequency band execution vector, and sort the noise reduction execution items to generate a noise reduction execution scheme.

[0029] The noise reduction processing module is used to locate the corresponding target processing segment from the mixed audio signal for each noise reduction execution item in the noise reduction execution scheme, convert the target processing segment into a short-time spectral representation and obtain the original spectral coefficients, generate transition coefficients according to the change type and the transition mode between short-time segments, calculate the actual spectral retention coefficients based on the transition coefficients, apply the actual spectral retention coefficients to the original spectral coefficients corresponding to the target processing segment to generate corrected spectral coefficients, summarize the corrected spectral coefficients corresponding to all noise reduction execution items and convert them into time-domain audio, overlap and splice the time-domain audio according to the original time order of the short-time segments, and output the noise-reduced output signal.

[0030] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0031] The beneficial effects of this invention are as follows:

[0032] This application first divides the mixed audio signal collected by the communication terminal into continuous short-time segments, and combines normalized short-time energy, adjacent change intensity, and a lightweight time-frequency feature extraction network to form a noise state feature set arranged in time, so that each segment contains both the spectral shape and the degree of change between adjacent segments. Then, based on the noise state feature set, a noise state change description is generated, and different engineering changes such as continuous high noise, wind noise enhancement, short-time impact, and noise shape switching are distinguished by comprehensive scoring, so that the algorithm can clearly identify the location, scope, type, and intensity of the change. On this basis, the noise state change description is transformed into a noise reduction execution scheme, and the target processing segment, normalized processing intensity, frequency band execution vector, and inter-segment transition mode are generated according to the change start segment, end segment, change type, and change intensity, so that different noise types correspond to different processing rhythms and frequency band allocations.

[0033] Finally, the target audio segment is processed in the short-time frequency domain according to the noise reduction execution plan. The spectral coefficients of the corresponding frequency band are processed using the actual spectral preservation coefficients, and the noise-reduced output signal is obtained through time-domain reconstruction. The core improvement of the above scheme lies in connecting the relatively scattered feature extraction, change judgment, and frequency band processing actions in existing adaptive noise reduction into a continuous decision chain oriented towards professional communication conditions. This allows the noise state information formed in the previous stage to be transformed into executable processing parameters in the subsequent stage, thereby implementing the identification results of noise state switching into actual noise reduction actions. In this way, the equipment can generate differentiated processing schemes based on the time range, type, and intensity of noise changes in typical field environments such as wind noise intrusion, mechanical start-up, broadband impact, and multi-source mixed noise. It maintains the continuity and clarity of communication voice while suppressing environmental noise, making it more suitable for deployment in tactical communication headsets, fire headsets, aviation headsets, and other professional communication terminals. Attached Figure Description

[0034] Figure 1 This is a flowchart of the adaptive noise reduction method in a specific embodiment of the present invention;

[0035] Figure 2 This is a block diagram of the adaptive noise reduction system in a specific embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] refer to Figure 1 As shown, one embodiment of this application proposes an adaptive noise reduction method, including:

[0038] S1: Acquire the mixed audio signal and divide it into short-time segments. After normalizing the short-time segments, calculate the normalized short-time energy of each short-time segment. Based on the normalized short-time energy, calculate the normalized change intensity between adjacent short-time segments. Convert the normalized short-time segments into a spectrum. Input the spectrum, normalized short-time energy, and normalized change intensity into a lightweight time-frequency feature extraction network. The lightweight time-frequency feature extraction network outputs a noise state feature set, specifically including:

[0039] The input for this step is a mixed audio signal. The signal is acquired by a pickup device on a professional communication audio device, which can be an environmental microphone on the outside of a tactical communication headset, a stick microphone near the mouth of a fire-fighting headset, a microphone on a police walkie-talkie handle, or a microphone for aviation headsets. The mixed audio signal... This represents the audio input during the current communication process, typically including components such as the wearer's voice, breathing sounds, wind noise, vehicle sounds, mechanical sounds, alarm sounds, and collision sounds. For example, when a firefighter wears a communication headset into a pumping station, the microphone collects... The noise may simultaneously include firefighters' voices, the sound of airflow from breathing apparatus, continuous low-frequency noise from pumps, and short, sharp impact sounds from metal tools colliding. The device receives this noise in its local processor. Subsequently, the continuous audio stream is divided into short segments arranged in chronological order. Each short segment retains its sequential relationship within the original audio stream, allowing subsequent processing to observe the change in noise state from one segment to the next. These short segments are continuously divided according to a fixed frame length and frame shift preset by the device. When the remaining sampling points at the end of the audio stream are insufficient for a complete short segment, zero-value padding is used to complete the segment before it continues to participate in subsequent processing. For example, wind noise typically increases gradually in continuous segments when turning against the wind, mechanical equipment transitions from low-amplitude disturbances to a stable roar when starting up, and metallic collision sounds experience a broadband surge within a short period. Therefore, continuous short segments more closely resemble the actual evolution of noise over time in professional communication environments.

[0040] Before each short segment is processed for feature extraction, amplitude normalization is performed to compress the amplitude differences caused by different microphone sensitivities, wearing distances, and speech loudness into a uniform range. The normalized segment amplitude is denoted as... ,in Indicates the time sequence number of a short segment. This indicates the sampling point number within the short segment. If the peak amplitude of a short segment is lower than the device's preset mute threshold, the normalized result of that short segment is treated as a zero vector, but its temporal order is still preserved. Subsequently, the normalized short-time energy of each short segment is calculated. This calculation is derived from the classic definition of short-time energy in digital signal processing, which uses the squared average of the sampled amplitudes within a segment to represent the overall acoustic intensity of that segment. In this application, the normalized segment amplitudes are used in the calculation; therefore, the calculation results are used to describe relative strength, enabling audio segments from different device gain conditions to enter the same subsequent judgment process.

[0041] ;

[0042] in, Indicates the first The normalized short-time energy of a short time segment is used to describe the overall intensity of the sound within that segment. Indicates mixed audio signal After being divided and normalized, the first The first short segment Each sample value is derived from the microphone signal after local normalization. This indicates the number of sampling points involved in the calculation within that short time segment. Because... It is the normalized amplitude ratio, obtained by averaging the squares. It is still a normalized proportional quantity, and the left and right sides of the formula express the same type of relative strength.

[0043] In obtaining Then, the normalized change intensity between adjacent short time segments is calculated. This is used to represent the degree of change of the current short segment relative to the previous short segment. This calculation originates from the concept of relative rate of change common in numerical analysis and signal detection, which uses the ratio between the difference between adjacent values ​​and their combined magnitude to represent the magnitude of change. This application adds a fixed smoothing term to this concept, making the calculation more stable near extremely low-energy segments and making the change intensity suitable for comparing consecutive segments in professional communications. A continuously operating wind turbine may maintain high energy for a long time, but the change between adjacent segments is small; when wind noise suddenly enters or collision sounds occur, the degree of change between the current segment and the previous segment will be more obvious. Able to supplement State transition information that cannot be expressed alone. For the first short segment, since there is no previous short segment, the terminal initializes its normalized change intensity to zero and performs adjacent segment comparison starting from the second short segment.

[0044] ;

[0045] in, Indicates the first The normalized change intensity of each short segment relative to the previous short segment is used to describe the degree of fluctuation in the noise state. This represents the normalized short-time energy of the current short-time segment; Represents the normalized short-time energy of the preceding short-time segment; fixed smoothing term. This is used to keep the denominator stable. Both the numerator and denominator in this formula are composed of normalized short-time energies, and the calculation results... This indicates the relative degree of change. Taking a simplified segment as an example, if the normalized sample value of a certain segment is... , , , Then the fragment for If the previous segment for ,but for Divide by , received In a stable pump operation scenario, adjacent segments... Typically smaller; in tool collision or strong wind intrusion scenes, adjacent segments... It will increase significantly, so this value can provide a direct basis for describing subsequent noise state changes.

[0046] The normalized short-time energy of each short-time segment is obtained. and adjacent change intensity Then, this step continues to generate time-frequency representations for the same short time segment. Specifically, the terminal converts the normalized short time segment into a spectrogram format and then... and A lightweight time-frequency feature extraction network is used for input. If a short time segment cannot form a valid spectrum due to missing sampling or an abnormal input buffer, the corresponding placeholder spectrum is generated directly using the spectrum size configuration of the previous valid short time segment, and an abnormal input flag is recorded on the short time segment. This network consists of an input sorting layer, two sets of convolutional processing layers, a set of gated recurrent processing layers, and an output mapping layer. The input sorting layer receives the spectrum of each short time segment, the segment's... and the corresponding The network is processed in chronological order, with the first convolutional layer extracting local frequency band textures, such as the expansion pattern of continuous wind noise in the low-frequency region and the stable stripe pattern of mechanical noise in some frequency bands. The second convolutional layer expands the frequency band observation range to identify broadband impact sound, multi-source mixed noise, and inter-band energy migration. The gated loop processing layer reads the convolutional outputs of consecutive segments in chronological order, combining the state representation of the current segment with the changing trend of the previous segment. The output mapping layer organizes the network results corresponding to each segment into a fixed-length noise state feature vector of the segment. The parameters of this network are set before the device leaves the factory based on samples of typical professional communication scenarios. Typical samples can come from fire fan environments, police vehicle patrol environments, aviation ground support environments, mechanical equipment room environments, and outdoor strong wind environments. After deployment to the terminal, the device will process the currently collected mixed audio signals. The noise state feature set is obtained by sequentially performing fragmentation, normalization, short-time energy calculation, change intensity calculation, and lightweight network processing. .

[0047] The output of this step is a noise state feature set. The aforementioned It is formed by arranging the noise state feature vectors of segments corresponding to consecutive short-time segments in chronological order. The noise state feature vector of each segment originates from the spectral information and normalized short-time energy of the same segment. Adjacent change intensity And the output of the lightweight time-frequency feature extraction network. This is how it is obtained. It simultaneously preserves the spectral morphology within a segment, the relative intensity within a segment, and the degree of change between segments, thus expressing the current state of the noise and how that state changes within consecutive segments. Through the above processing, the mixed audio signal... Transformed into a noise state feature set for professional communication scenarios This feature set can reflect the characteristics of strong noise, complex sources, and rapid state changes in fire fighting, police, aviation, and special operation environments, and serves as the input for the next step of generating a description of noise state changes.

[0048] S2: Calculate a comprehensive score for short segments in the noise state feature set. Based on the comprehensive score, normalized change intensity, and the noise state feature vector corresponding to the short segment, determine the starting segment of the change, the range of influence of the change, and the type of change, and construct a description of the noise state change, specifically including:

[0049] The input for this step is the noise state feature set output by S1. The aforementioned Arranged in chronological order of short-time segments, each short-time segment corresponds to a segment noise state feature vector, and the noise state feature vector of the same segment has been fused with the spectral morphology and normalized short-time energy of that segment. Adjacent change intensity And the state representation of a lightweight time-frequency feature extraction network. This step will... The segment noise state feature vector corresponding to each short time segment is denoted as . Let the segment noise state feature vector corresponding to the previous short segment be denoted as Both are directly taken from And read according to the time sequence in which S1 was formed. The core problem to be solved in this step is to... The acoustic state representation of continuous segments is transformed into a description of noise state changes. This allows subsequent steps to clearly identify which segments the noise change occurs in, which segments it affects, the type of engineering involved, and the intensity of the change. For professional communication scenarios, this transformation is crucial because background noise from sources such as fire pumps, police vehicles, and aviation ground equipment may remain at a high intensity for extended periods, but as long as its spectral characteristics and changes in adjacent segments remain stable, it remains in a continuous state. Conversely, events like turning against the wind, equipment startup, and tool collisions, although shorter in duration, create significant state transitions between adjacent segments, requiring analysis to inform subsequent noise reduction strategies.

[0050] This step first involves... The state differences between adjacent segments are comprehensively scored. The initial sources for this score include three classic computational ideas: one is the detection of energy changes between adjacent frames in digital signal processing, used to describe the acoustic fluctuations between segments; this part directly uses the output of S1. Secondly, it utilizes the cosine similarity of vectors in linear algebra to describe the closeness of two noise state feature vectors in terms of morphological direction. This application subtracts the cosine similarity from the first value to transform it into a difference in the morphological shape of the noise state feature vectors. Thirdly, it employs the first-order difference concept in numerical computation to describe the rate of change of adjacent change intensities. This application adopts... This indicates the degree of sudden increase in the rate of change. The above three types of information are combined into a comprehensive score. ,in Reflecting sound fluctuations, the cosine difference indicates whether noise texture and spectral states are switching. This reflects the accelerated changes commonly seen when wind noise, impact noise, or mechanical startup suddenly enters. This combination method allows continuous high noise, progressively increasing noise, and short-term impact noise to be distinguished under the same scoring rules; when the noise state feature vectors of two adjacent segments both correspond to silent segments or their magnitudes are both below a preset vector threshold, the cosine difference term is treated as zero to avoid misjudging silent segments as state transitions.

[0051] ;

[0052] in, Indicates the first The overall score of a short segment is used to describe the degree of change in the noise state of the segment relative to the previous segment; This represents the preset weighting coefficient, determined during equipment factory calibration based on typical scenario samples such as fire-fighting fans, police vehicles, aviation ground equipment, and strong outdoor winds. It is used to adjust the intensity of adjacent changes. The proportion between the morphological differences of the noise state feature vector; Indicates the output of S1 of the first... The normalized change intensity of each short segment relative to the previous short segment; Indicates the first The normalized change intensity of a short time segment relative to its previous short time segment; The preset weights for the rate of change are determined by the equipment during calibration based on samples of sudden impact sound and wind noise. Representing the noise state feature set The Middle The first short-time segment noise state feature vector One component; Indicates the first The first short-time segment noise state feature vector One component; The number of components contained in the feature vector of a segment noise state; fixed smoothing term. Used to keep the denominator calculation stable. In the formula cosine difference term and All are derived from the normalized state variables obtained from S1, and they all express the relative degree of change between adjacent segments. The result after merging is... It is still used to represent the relative strength of state transitions.

[0053] The practical application of this calculation can be illustrated using a fire pump room scenario. When the pump is running stably, the overall sound of consecutive segments is relatively strong, but the states between segments are similar; assuming the current segment's... for The previous segment for The cosine similarity of adjacent noise state feature vectors is Equipment calibration results for , for ,but for Multiply , plus Multiply In addition Multiply The result is approximately This indicates that the segment is in a stable, continuous high-noise state. If wind noise suddenly enters after a firefighter approaches the exhaust vent, the current segment's... Rise to The previous segment for The cosine similarity of adjacent noise state feature vectors decreases to Under the same calibration parameters, Approximately This indicates that a significant noise state transition has occurred in the segment. In this example, the stable background noise from the pump receives a lower score due to the similar shape of the noise state feature vector and the low rate of change; the sudden wind noise receives a higher score due to the simultaneous increase in energy fluctuations, the difference in the shape of the noise state feature vector, and the rate of change, which meets the engineering requirement of prioritizing the identification of noise state transitions in professional communication scenarios.

[0054] In obtaining each short time segment in succession Then, the terminal forms the natural fluctuation level of the current environment based on previous stable segments. This natural fluctuation level is determined by analyzing recent segments. The noise level is obtained by sorting and taking the median level, which represents the normal fluctuation of the current background noise. The recent segment is the most recent preset number of valid short-time segments. If the number of valid short-time segments obtained is less than the preset number, the natural fluctuation level is calculated based on the currently obtained valid short-time segments. For example, when a police vehicle is traveling at a constant speed, engine and road noise will form a relatively stable natural fluctuation level; when talking to someone near aviation ground support equipment, the equipment noise will also form a relatively stable natural fluctuation level. The terminal will then sort the current segment and its subsequent adjacent segments. Compared to this natural fluctuation level, when multiple consecutive segments The feature vector of the corresponding segment noise state is higher than the sum of the natural fluctuation level and the preset rise threshold. The direction of morphological change and When the trend of change meets the preset consistency judgment rule, the segment that first meets the condition is determined as the starting segment of the change. Subsequently, the terminal continues to read subsequent segments in chronological order. and Fragments exhibiting persistent change characteristics are grouped into the same scope of change influence; when When the current environmental fluctuation level returns to near and the feature vector of the segment noise state stabilizes again, the terminal determines the end position of the range of influence of the change; if the ranges of influence of the change are separated by only a few short segments and the change type is the same, they are merged into the same range of influence of the change.

[0055] After determining the initial segment of the change and the scope of its impact, the terminal further organizes the change types. If the initial segment of the change is near... It rose rapidly and then fell back in a short period of time. If it manifests as broadband texture enhancement, it is then processed into a short-term impact-type change, suitable for scenarios such as metal collisions, tools falling, and near-field popping sounds; if It gradually increases in multiple consecutive segments, and The mid-to-low frequency extension pattern is enhanced, while If it shows a continuous upward trend, it is classified as a wind noise enhancement type change, suitable for scenarios such as patrol vehicles turning into the wind, firefighters approaching ventilation vents, and strong airflow on the ground during aviation; if The changes were relatively gradual, but If a significant shift occurs in the spectral texture, it is categorized as a noise pattern switching change, applicable to scenarios such as transitioning from low-frequency mechanical background to alarm sounds, or from vehicle noise to crowd noise. If the same change's influence area simultaneously meets multiple change conditions, the change type is determined according to the priority order of short-term impact change, wind noise enhancement change, and noise pattern switching change; if it still cannot be clearly categorized, it is treated as a noise pattern switching change. This categorization process directly uses... The continuous segment noise state feature vector and S1 obtained , This allows the same set of noise state characteristics to be further transformed into a change description with engineering implications.

[0056] The output of this step is a description of noise state changes. The aforementioned It consists of change descriptions arranged chronologically. Each change description includes the starting segment of the change, the scope of the change's impact, the type of change, and the intensity of the change. The starting segment of the change is determined by a comprehensive score. The location of the first significant increase was determined; the range of influence of the change was determined by subsequent consecutive segments. The continuation status and noise state feature vector The morphological continuity is jointly determined; the type of change is determined by the area near the point of change. form, Trends and The fluctuations are combined to obtain the result; the intensity of the change is determined by the range of influence of that change. Horizontal formation. Taking police motorcycle patrol as an example, when the vehicle is traveling at a constant speed, The corresponding record is a continuation of a stable driving background; when the vehicle enters a windward section, the continuous segments... , And low-frequency extended forms rise in tandem. A wind noise enhancement variation term is formed; when a brief metallic collision occurs near the headphones... The area rapidly increases in size at the collision point and the broadband texture becomes prominent. This generates short-term impulse-type change terms. Through the above processing, the noise state feature set output by S1 is obtained. It is transformed into a description of noise state changes that can specify the location, range, type, and intensity of the change. This serves as the direct input for the next step of generating a noise reduction execution plan for the target time period.

[0057] S3: According to a preset execution mapping table, calculate the normalized processing intensity of the state change description, calculate the frequency band execution vector of the change type based on the normalized processing intensity, construct noise reduction execution items based on the state change description, normalized processing intensity, and frequency band execution vector, and sort the noise reduction execution items to generate a noise reduction execution scheme, specifically including:

[0058] The input for this step is the noise state change description of the S2 output. The aforementioned It consists of change descriptions arranged in chronological order, with each change description corresponding to a identified noise state change process. The first... Each change description item is denoted as , which contains the starting fragment of the change End of change segment Types of change and intensity of change .in, and Overall score from S2 The location of the rise and fall of the noise level indicates the range of influence of the noise change in a continuous short time segment. The morphology of the segment noise state feature vector near the change point in S2, and the normalized short-time energy Trend and strength of adjacent changes The fluctuations are combined to indicate whether the change belongs to wind noise enhancement type, short-term impact type, or noise pattern switching type. Take the comprehensive score within the scope of the impact of this change. The maximum value is used to represent the highest intensity of change achieved within the corresponding area of ​​influence. This step reads data in chronological order. Each of them and will Transform it into the processing starting point in the execution plan, Transform it into the processing endpoint in the execution plan, This is converted into the processing type and frequency band configuration in the execution plan. This is transformed into processing intensity within the execution plan. Therefore, the description of the output change in S2 is further refined into a noise reduction execution plan that S4 can directly execute. .

[0059] In professional communication scenarios, the generation of noise reduction execution schemes needs to consider the intensity, type, and duration of changes simultaneously. For example, continuous low-frequency rumbling in a fire pump room needs to be stably suppressed; the increased wind noise when a police motorcycle turns into the wind requires smoothing of low-frequency disturbances within continuous segments; short-term impacts from metal tool collisions need to be quickly suppressed and recovered within a short time; and the jet noise from equipment in aviation ground support may manifest as continuous disturbances superimposed with local broadband variations. Based on this, an execution mapping table is pre-set in the terminal firmware, and the execution mapping table is categorized according to the type of change. Record the corresponding type correction value, frequency band processing priority, and voice preservation configuration, with the frequency band processing priority and voice preservation configuration written according to a normalized ratio; if a certain change type does not match the execution mapping table, the preset general noise configuration is invoked. The terminal reads... Then, first according to Retrieve the type correction value from the execution mapping table. Combined with the intensity of change Calculate the normalization intensity corresponding to the variation description item based on the variation range length and equipment calibration coefficient. This calculation originates from the linear mapping and saturation limiting concepts in adaptive gain control. Its basic logic is to map the input change to a control strength and maintain the execution strength within a stable range using upper and lower boundaries. In this application, the basic linear mapping term is... This indicates that the impact of the change type on the executed action is determined by... This indicates that the impact of the duration of change on the executed action is represented by the normalized duration term, and the three are combined before entering the amplitude limiting process.

[0060] ;

[0061] in, Indicates the first The normalization processing intensity corresponding to each change description item is written into the noise reduction execution scheme. ; This represents the preset intensity mapping coefficient, which is determined by the equipment during factory calibration based on samples from fire-fighting fans, police vehicles, aviation ground equipment, and strong outdoor winds. Indicates the first Each change description item The intensity of change recorded in S2 is determined by the comprehensive score within the corresponding range of influence of the change. The results were obtained through sorting. Indicates the type of change The type correction values ​​obtained from the execution mapping table are as follows: short-term impact type changes correspond to higher correction values, wind noise enhancement type changes correspond to medium correction values, and noise pattern switching type changes correspond to gentle correction values. The weights of the terms representing the duration of change are obtained from the equipment's factory calibration. Indicates the starting segment of the change. Indicates the end of a change segment; both originate from... ; This represents the stable configuration parameter of the persistence term, used to control the rate at which the length of the variation range affects the processing intensity. In the formula... , , Continuity of change and All are normalized proportions. and Used for proportional mapping and proportional weighting, both sides of the formula represent the intensity of the normalization process. This formula consists of the basic linear mapping term. As it unfolds, it increases This allows the same intensity of change to be applied differently depending on the type of noise. Adding a continuous term allows for more thorough handling of changes with a longer duration, such as continuous wind noise and equipment startup. Finally, an executable and stable intensity is formed through the upper and lower boundaries.

[0062] Taking a police motorcycle patrol scene facing the wind as an example, the change description item generated by S2 This indicates a wind noise enhancement change, starting segment of the change. For the first A short segment, the end of the change segment For the first A short time segment, the intensity of change for Equipment calibration results for Wind noise enhancement type change corresponding to for , for , for The length of the range of change at this time is ,Right now A short time segment, with a duration term of... Divide by , approximately The candidate value for processing intensity is... Multiply , plus In addition Multiply , received This value is within the execution boundary, therefore Pick In this scenario, the terminal will... To the A short time segment is used as the main target processing range, and a smooth transition is configured according to the continuous characteristics of wind noise enhancement changes, so that the processing action covers the main stage of wind noise enhancement and the adjacent transition stage. If Indicates a short-term shock-type change. Taking a higher value usually results in a shorter range of variation, and the calculated value is... Mainly composed of and The decision is made to allow for faster suppression actions in metal collision segments and their immediate subsequent segments.

[0063] In obtaining Then, the terminal continues to adjust according to the type of change. Generate frequency band execution vector This calculation originates from the classic ideas in band-weighted filtering and spectral gain control, which involves setting different execution weights for different frequency bands so that the noise reduction process can perform differentiated actions according to the noise distribution across the frequency bands. In this application, the execution mapping table is for each type of variation. Configure frequency band processing priority vector and speech-preserving vectors .in, This indicates the noise processing priority of the change type in each frequency band. For example, wind noise enhancement type changes have a higher processing priority in the low-frequency disturbance region, short-time impulse type changes have a higher processing priority in multiple frequency bands at the same time, and noise pattern switching type changes are reassigned in a medium intensity frequency band according to the spectrum shift configuration before and after the change. This indicates the voice preservation configuration for each frequency band under this change type, used to create smoother processing in frequency bands related to command clarity. The terminal will... , and By combining these elements, the frequency band execution vector corresponding to the change description term is obtained. and write it .

[0064] ;

[0065] in, Indicates the first The frequency band execution vector corresponding to each change description item is written into the noise reduction execution scheme. ; This represents the normalized intensity obtained from the aforementioned calculation; Indicates the type of change The frequency band processing priority vector is retrieved from the execution mapping table; Indicates the type of change The speech-preserving vector is retrieved from the execution mapping table; This indicates that the corresponding positions are multiplied, which is used to ensure that the execution weight of each frequency band is simultaneously constrained by noise processing priority and speech preservation configuration. In the formula... , , and All are normalized proportional values, and the calculation results represent the set of execution weights corresponding to each frequency band. The logical relationship between this formula and the previous formula is as follows: First from Calculate the overall processing intensity based on the intensity, type, and range of change. Then by Generate by combining the frequency band configuration corresponding to the change type This transforms a change description into an execution item that has both overall strength and bandwidth allocation.

[0066] Continuing with the above scenario of police motorcycles facing the wind, the wind noise enhancement change corresponds to... It can be configured as This indicates that low-frequency disturbance regions and adjacent frequency bands have higher processing priority; the corresponding It can be configured as This indicates that the closer to the frequency band related to speech clarity, the higher the reserved configuration. When for hour, The execution weights of the four frequency bands are approximately Multiply Multiply by , Multiply Multiply by , Multiply Multiply by , Multiply Multiply by , received This means that in the first... To the Within a short time segment and its smooth transition range, the terminal performs stronger processing on low-frequency wind noise areas and smoother processing on frequency bands related to speech clarity. If For short-term, impactful changes, the execution mapping table will be configured to be more balanced. This allows multiple frequency bands to obtain higher execution weights in a short period of time; if This is a noise morphology switching change. and This will allow the frequency band processing to gradually transition to a new noise state within the changing range.

[0067] The terminal is for each generate and Then, the target processing segment, processing type, processing intensity, frequency band execution vector, and inter-segment transition method are organized into a single noise reduction execution item. The target processing segment is composed of... and Determine, and according to The engineering characteristics configuration transition range is as follows: wind noise enhancement changes adopt a smooth backward continuation; short-term impact changes adopt a short-range rapid recovery; noise pattern switching changes cover the affected area and adopt a segmented transition. The transition method between segments is based on... The magnitude and temporal relationship between adjacent execution items are determined. High-intensity impact-type execution items use faster entry and faster recovery; continuous wind noise execution items use smooth entry and smooth exit; and morphological switching-type execution items use a gradual transition covering the range of change. If two execution items overlap in time, the execution item with the higher intensity is normalized and designated as the primary execution item. The lower intensity execution item is only retained as an auxiliary item in a frequency band that does not conflict with the primary execution item. If the two change types are consistent and their times are continuous or overlapping, they are directly merged into one execution item. Taking a fire scene as an example, when a tool collision occurs against a background of pump noise... The process may simultaneously contain both continuous background changes and short-term impact changes; the terminal organizes the two execution items in chronological order, and in the collision segment, the higher value corresponding to the short-term impact change is used. and broadband The low-frequency processing arrangement corresponding to the pump background is retained in the continuous segments after the collision, so that... It is continuous in time and has a clear division of labor in frequency band.

[0068] The output of this step is the noise reduction execution plan. The aforementioned It consists of noise reduction execution items arranged in chronological order, with each execution item consisting of a change description item. Generate, and include at least the target processing fragment, processing type, and normalized processing intensity. Frequency band execution vector Transition methods between segments. Target processing segments are composed of... The starting segment of the change and the end of the change segment Generation; Processing type is determined by Types of change in Generation; Normalization of intensity From the intensity of change Type correction value Generated together with the duration of change; frequency band execution vector Depend on Frequency band processing priority vector and speech-preserving vectors Generation; the transition method between segments is determined by The time relationship with adjacent execution items is obtained by sorting out the data. Through the above processing, the noise state change description of S2 output is obtained. Transformed into a noise reduction execution scheme that can be directly invoked by S4 This allows the adaptive noise reduction algorithm to move from identifying changes in noise state to the specific processing arrangements for the target segment.

[0069] S4: For each noise reduction execution item in the noise reduction execution scheme, locate the corresponding target processing segment from the mixed audio signal, convert the target processing segment into a short-time spectral representation and obtain the original spectral coefficients, generate transition coefficients based on the change type and the transition mode between short-time segments, calculate the actual spectral retention coefficients based on the transition coefficients, apply the actual spectral retention coefficients to the original spectral coefficients corresponding to the target processing segment to generate corrected spectral coefficients, summarize the corrected spectral coefficients corresponding to all noise reduction execution items and convert them into time-domain audio, overlap and splice the time-domain audio according to the original time order of the short-time segments, and output the noise-reduced output signal, specifically including:

[0070] The input for this step is the noise reduction execution scheme output by S3. and call the mixed audio signal Zhongyou The specified target processing segment is used as the object to be processed. It consists of noise reduction execution items arranged in chronological order. Each noise reduction execution item corresponds to a target processing procedure and includes the target processing segment, processing type, and normalization processing intensity. Frequency band execution vector Transition methods between segments. Terminal reading. Then, according to the time sequence of each noise reduction execution item, the mixed audio signal is extracted from the local audio buffer. The corresponding target processing segment is located and converted into a short-time spectral representation; the short-time spectral representation uses the same short-time segment division method as S1. If the target processing segment specified by an execution item exceeds the current valid range of the audio buffer, it is truncated to the nearest available segment according to the valid range of the buffer. This conversion adopts the short-time frequency domain processing method commonly used in digital audio processing, expanding the target processing segment into several short-time processing units according to its original short-time segment sequence, and obtaining the spectral coefficients of each short-time processing unit in each frequency band. Within the noise reduction execution item, the first The first short-time processing unit, the first The original spectral coefficients corresponding to each frequency band are denoted as The spectral coefficients are composed of mixed audio signals. The target processing segment specified by the execution item is obtained through short-time frequency domain transformation. Subsequently, the terminal reads the processing type and frequency band execution vector from the same execution item. The transition method between segments, where the processing type determines the shape of the transition curve. The execution weight of each frequency band is determined, and the transition method between segments determines the connection between adjacent short-time processing units; among them, The intensity has been normalized in S3. Frequency band processing priority vector and speech-preserving vectors Together they are generated. Therefore, the output of S3 is... All kinds of information are then incorporated into the actual audio processing.

[0071] In practice, the terminal generates transition coefficients based on the processing type and the transition method between segments. Regarding wind noise enhancement variations, The frequency is gradually increased at the beginning of the target processing segment and gradually decreased at the end of the segment, allowing for a smooth entry and exit of low-frequency disturbance processing; for short-term impact-type changes, The noise level rises rapidly near the initial segment and then drops rapidly in the immediate following segment, allowing for concentrated processing of short-term noises such as metal collisions, tool drops, and near-field impact sounds; for noise pattern switching changes... The frequency band processing is gradually changed within the range of influence of the change, so that the frequency band processing can smoothly transition from the original noise state to the new noise state. Generates and constrains data to a normalized ratio between zero and one; when the target processing range only covers a single short-term processing unit, a symmetrical, fast-entry and fast-exit single-unit transition method is adopted. Thus, Processing type, frequency band execution vector Together with the transition method between segments, the processing actions within the target segment are determined, so that the processing actions correspond to the noise type and maintain the continuity between adjacent segments.

[0072] The terminal then calculates the actual spectral retention coefficient. This calculation originates from the classic concept of frequency domain spectral gain processing, which involves scaling the frequency domain spectral coefficients using a proportional gain coefficient to alter the energy of a specific frequency band. In this application, the frequency domain spectral gain is expressed by a "retention coefficient." The smaller the retention coefficient, the stronger the noise component of the corresponding frequency band is processed; the closer the retention coefficient is to one, the more of the corresponding frequency band is retained. This classic concept is improved in this step by a noise reduction execution scheme. Direct drive: Band execution vector Decision No. Processing weights and transition coefficients for each frequency band Decision No. Each short-time processing unit is in the entry, stabilization, or exit phase, thus forming the actual spectral retention coefficients that simultaneously exhibit band differences and time transition characteristics.

[0073] ;

[0074] in, Indicates the first The first noise reduction execution item The first short-time processing unit, the first The actual spectral retention coefficients of each frequency band; Indicates by Transition coefficients generated by processing type and inter-segment transition method; Indicates S3 output and writing The The noise reduction execution item is in the first The execution weights on each frequency band, which have been normalized by the intensity in S3. Frequency band processing priority vector and speech-preserving vectors Calculated jointly. Because... and Both are normalized proportions; their product still represents the normalized proportion. This indicates the retention ratio of the corresponding frequency band spectral components.

[0075] Obtain the actual spectral retention coefficient Then, the terminal applies it to the original spectral coefficients corresponding to the target processing segment. The processed spectral coefficients are obtained. This calculation continues the classic form of frequency domain gain filtering, that is, scaling the original spectral coefficients using spectral preservation coefficients; the improvement in this step is that these spectral preservation coefficients are generated by S3. The decision allows for differentiated processing based on target segments, frequency bands, weights, and transition states. The previous formula is used to... The actual spectral retention coefficients for each frequency band and each short-time processing unit are generated. This formula is used to apply these retention coefficients to the corresponding spectral coefficients, and together they complete the process from the noise reduction implementation scheme to the actual audio signal change.

[0076] ;

[0077] in, Indicates the first The first noise reduction execution item The first short-time processing unit, the first Corrected spectral coefficients of each frequency band after noise reduction processing; This represents the aforementioned actual spectrum retention coefficient; express The target processing segment pointed to The first short-time processing unit, the first The original spectral coefficients of each frequency band. Because... For normalized proportions, and Maintain the same spectral coefficient expression.

[0078] Continuing the police motorcycle windswept scene from Season 3, the scene generated in Season 3... Specify the To the Each short time segment represents the processing range for wind noise enhancement targets, and a frequency band execution vector is obtained. Approximately When processing the middle of the target segment, the transition coefficient... Pick The actual spectral retention coefficient of the first frequency band is... for minus ,get The actual spectral retention coefficient of the fourth frequency band for minus ,get If the original spectral coefficients of the first frequency band are Spectral coefficients after processing For the original spectral coefficients If the original spectral coefficients of the fourth frequency band are... Spectral coefficients after processing For the original spectral coefficients This processing effectively suppresses low-frequency wind noise while maintaining a high retention rate in higher frequency bands related to speech intelligibility. If the same execution item is in the entry phase of the target segment, For example, take Then the first frequency band for minus Multiply , received This allows the processing action to smoothly transition into the target segment. For short-duration, impactful changes, The transition method in the middle will make The impact segment rises rapidly and then falls rapidly, thus creating strong processing within a short period of time during the impact, restoring the naturalness of the speech in subsequent communication segments.

[0079] Terminal to The above process is repeated for each noise reduction execution item, and the processed spectral coefficients are reconstructed according to the temporal order of each execution item. For temporally adjacent execution items, the terminal uses their respective transition coefficients. The handover is completed, allowing the processing of the previous execution item to gradually exit and the processing of the next execution item to gradually enter. For execution items that overlap in time, the terminal uses the main execution item and auxiliary frequency band configuration after S3 processing to ensure consistent frequency band processing within the same target segment. If no noise reduction execution item is generated in the current analysis window, the current short-time spectral representation is maintained unchanged, and time-domain reconstruction begins. Taking a fire scene as an example, the low-frequency processing execution item corresponding to the pump background sound and the short-time impact execution item corresponding to the tool collision may be temporally adjacent. The terminal uses the broadband processing weight corresponding to the short-time impact change in the collision segment, and continues to use the low-frequency processing weight corresponding to the pump background sound in subsequent segments. After all target processing segments have completed spectral coefficient processing, the terminal will process the spectral coefficients. The audio is converted back to the time domain and then overlapped and spliced ​​according to the original time sequence of the short segments to form a noise-reduced output signal. .

[0080] The output of this step is a noise-reduced output signal. The aforementioned Depend on The target processing segment pointed to by each noise reduction execution item in the process is formed after frequency band execution, transition splicing, and temporal reconstruction. Through this processing, the target processing segment, processing type, and normalized processing intensity output by S3 are determined. Frequency band execution vector The transition methods between segments are all transformed into actual audio processing actions, ultimately realizing the changes such as wind noise enhancement, short-term impact, and noise pattern switching in professional communication scenarios into corresponding noise reduction outputs.

[0081] refer to Figure 2 As shown, another aspect of this application embodiment proposes an adaptive noise reduction system, including:

[0082] The feature extraction module is used to acquire mixed audio signals and divide them into short time segments. After normalizing the short time segments, it calculates the normalized short-time energy of the short time segments, calculates the normalized change intensity between adjacent short time segments based on the normalized short-time energy, converts the normalized short time segments into spectrograms, and inputs the spectrograms, normalized short-time energy, and normalized change intensity into a lightweight time-frequency feature extraction network. The lightweight time-frequency feature extraction network outputs a noise state feature set.

[0083] The state change description module is used to calculate a comprehensive score for short segments in the noise state feature set, and to determine the starting segment, the range of influence and the type of change based on the comprehensive score, the normalized change intensity and the noise state feature vector corresponding to the short segment, and to construct a state change description of the noise.

[0084] The noise reduction scheme generation module is used to calculate the normalized processing intensity of the state change description according to a preset execution mapping table, calculate the frequency band execution vector of the change type in combination with the normalized processing intensity, construct noise reduction execution items in combination with the state change description, normalized processing intensity and frequency band execution vector, and sort the noise reduction execution items to generate a noise reduction execution scheme.

[0085] The noise reduction processing module is used to locate the corresponding target processing segment from the mixed audio signal for each noise reduction execution item in the noise reduction execution scheme, convert the target processing segment into a short-time spectral representation and obtain the original spectral coefficients, generate transition coefficients according to the change type and the transition mode between short-time segments, calculate the actual spectral retention coefficients based on the transition coefficients, apply the actual spectral retention coefficients to the original spectral coefficients corresponding to the target processing segment to generate corrected spectral coefficients, summarize the corrected spectral coefficients corresponding to all noise reduction execution items and convert them into time-domain audio, overlap and splice the time-domain audio according to the original time order of the short-time segments, and output the noise-reduced output signal.

[0086] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0087] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0088] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An adaptive noise reduction method, characterized in that, include: A mixed audio signal is acquired and divided into short time segments. After normalizing the short time segments, the normalized short-time energy of the short time segments is calculated. Based on the normalized short-time energy, the normalized change intensity between adjacent short time segments is calculated. The normalized short time segments are converted into spectrograms. The spectrograms, normalized short-time energy, and normalized change intensity are input into a lightweight time-frequency feature extraction network. The lightweight time-frequency feature extraction network outputs a noise state feature set. A comprehensive score is calculated for short segments in the noise state feature set. Based on the comprehensive score, normalized change intensity, and noise state feature vector corresponding to the short segment, the starting segment of the change, the range of influence of the change, and the change type are determined, and a description of the noise state change is constructed. According to the preset execution mapping table, the normalized processing intensity of the state change description is calculated, and the frequency band execution vector of the change type is calculated in combination with the normalized processing intensity. The noise reduction execution items are constructed by combining the state change description, the normalized processing intensity and the frequency band execution vector, and the noise reduction execution items are sorted to generate a noise reduction execution scheme. For each noise reduction execution item in the noise reduction execution scheme, the corresponding target processing segment is located from the mixed audio signal, the target processing segment is converted into a short-time spectral representation and the original spectral coefficients are obtained, a transition coefficient is generated according to the change type and the transition mode between short-time segments, the actual spectral retention coefficient is calculated based on the transition coefficient, the actual spectral retention coefficient is applied to the original spectral coefficient corresponding to the target processing segment to generate a corrected spectral coefficient, the corrected spectral coefficients corresponding to all noise reduction execution items are summarized and converted into time-domain audio, the time-domain audio is overlapped and spliced ​​according to the original time order of the short-time segments, and the noise-reduced output signal is output.

2. The adaptive noise reduction method according to claim 1, characterized in that, The lightweight time-frequency feature extraction network includes an input sorting layer, a convolutional processing layer, a gated recurrent processing layer, and an output mapping layer. The sorting layer is used to receive the spectrum, normalized short-time energy, and normalized change intensity and arrange them in time order according to short-time segments. The gated recurrent processing layer is used to extract local frequency band texture and expand the observation range of the frequency band. The output mapping layer is used to sort the local frequency band texture and output it as a noise state feature set.

3. The adaptive noise reduction method according to claim 1, characterized in that, The determination of the starting segment, the range of influence, and the type of change based on the comprehensive score, the normalized change intensity, and the noise state feature vector corresponding to the short-time segment specifically includes: A preset number of short time segments are obtained, the comprehensive scores of the short time segments are sorted, and the median level is taken as the natural fluctuation level. When the comprehensive score of the short segment to be compared is higher than the sum of the natural fluctuation level and the preset rise threshold, and the change trend of the shape change direction and the normalized change intensity of the corresponding noise state feature vector meets the preset consistency judgment rule, the short segment is determined as the change start segment. Continue reading the comprehensive score and noise state feature vector of subsequent short segments in chronological order, and classify the short segments with continuous change features into the same range of change influence; When the difference between the comprehensive score of the short-term segment and the natural fluctuation level is less than the threshold, and the noise state feature vector is stable, the short-term segment is determined as the endpoint segment. The change type is obtained by combining the shape of the noise state feature vector around the change point, the change trend of the normalized short-time energy, and the fluctuation of the normalized change intensity. The noise state feature vector is obtained through the noise state feature set.

4. The adaptive noise reduction method according to claim 1, characterized in that, The step of converting the normalized short-time segment into a spectrum further includes: if a short-time segment cannot form a valid spectrum due to sampling loss or input buffer abnormality, then the spectrum size configuration of the previous short-time segment is used to generate a corresponding placeholder spectrum, and an abnormal input flag is recorded on the short-time segment.

5. The adaptive noise reduction method according to claim 1, characterized in that, The normalized short-time energy is calculated and generated using the sampled values ​​of the short-time segment and the number of sample points.

6. The adaptive noise reduction method according to claim 1, characterized in that, The comprehensive score is calculated and generated by normalizing the change intensity, preset weights, preset weighting coefficients, components of the noise state feature vector, and the number of components.

7. The adaptive noise reduction method according to claim 1, characterized in that, The step of calculating the normalization processing intensity of the state change description according to the preset execution mapping table specifically includes: The execution mapping table calls the corresponding type correction value, frequency band processing priority, and voice preservation configuration according to the change type. Based on the type correction value, frequency band processing priority vector, and voice preservation vector, the normalized processing intensity is calculated. The frequency band processing priority vector and voice preservation vector are both written according to the normalization ratio. If the change type does not match the execution mapping table, the preset general noise configuration is invoked.

8. The adaptive noise reduction method according to claim 1, characterized in that, The frequency band execution vector is calculated and generated by normalizing the processing intensity, the frequency band processing priority vector, and the speech preservation vector.

9. An adaptive noise reduction system, characterized in that, include: The feature extraction module is used to acquire mixed audio signals and divide them into short time segments. After normalizing the short time segments, it calculates the normalized short-time energy of the short time segments, calculates the normalized change intensity between adjacent short time segments based on the normalized short-time energy, converts the normalized short time segments into spectrograms, and inputs the spectrograms, normalized short-time energy, and normalized change intensity into a lightweight time-frequency feature extraction network. The lightweight time-frequency feature extraction network outputs a noise state feature set. The state change description module is used to calculate a comprehensive score for short segments in the noise state feature set, and to determine the starting segment, the range of influence and the type of change based on the comprehensive score, the normalized change intensity and the noise state feature vector corresponding to the short segment, and to construct a state change description of the noise. The noise reduction scheme generation module is used to calculate the normalized processing intensity of the state change description according to a preset execution mapping table, calculate the frequency band execution vector of the change type in combination with the normalized processing intensity, construct noise reduction execution items in combination with the state change description, normalized processing intensity and frequency band execution vector, and sort the noise reduction execution items to generate a noise reduction execution scheme. The noise reduction processing module is used to locate the corresponding target processing segment from the mixed audio signal for each noise reduction execution item in the noise reduction execution scheme, convert the target processing segment into a short-time spectral representation and obtain the original spectral coefficients, generate transition coefficients according to the change type and the transition mode between short-time segments, calculate the actual spectral retention coefficients based on the transition coefficients, apply the actual spectral retention coefficients to the original spectral coefficients corresponding to the target processing segment to generate corrected spectral coefficients, summarize the corrected spectral coefficients corresponding to all noise reduction execution items and convert them into time-domain audio, overlap and splice the time-domain audio according to the original time order of the short-time segments, and output the noise-reduced output signal.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

Images