Audio denoising method for a bluetooth communication earphone
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN HENGCLOUDS ECOLOGY TECH CO LTD
- Filing Date
- 2026-05-12
- Publication Date
- 2026-06-23
Smart Images
Figure CN122269187A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of audio processing technology, and in particular to an audio noise reduction method for Bluetooth communication headsets. Background Technology
[0002] Bluetooth headsets have become a common tool for mobile calls, remote conferencing, and online interaction. When users use headsets indoors near walls, in enclosed spaces like cars, or in environments with abundant reflective surfaces like streets, ambient noise enters the main and secondary microphones through multiple paths, significantly impacting the clarity of uplink speech. To suppress this background noise, a dual-microphone noise reduction architecture is commonly used. This architecture uses reference noise picked up by the secondary microphone to adaptively cancel noise from the main microphone signal, followed by post-processing spectral enhancement to further eliminate residual noise. Finally, the processed speech is encoded and transmitted.
[0003] In this architecture, the update of adaptive filter weights is typically gated by speech activity detection results to learn noise during periods without voice and maintain filter stability during periods with voice. However, when the acoustic characteristics of the user's location change, such as moving from an open space into a narrow corridor or near a large wall, the altered acoustic reflection conditions cause non-stationary fluctuations in the spatial coherence of the noise components from both microphones. In this case, speech activity detection based on fixed criteria and rigid thresholds is prone to misjudgment between pure noise and human voice regions, leading to incorrect filter coefficient updates during periods of voice or incorrect freezing during rapid changes in noise characteristics. This rigid binding-induced filter mismatch can partially cancel out human voice components or leave a large amount of abrupt noise, thus undermining the assumption of noise floor stability in post-spectral enhancement. This results in parameter estimation bias and gain oscillations in the speech enhancement stage, ultimately leading to cascading sound quality degradation in the output signal. Summary of the Invention
[0004] This application provides an audio noise reduction method for Bluetooth communication headsets, which solves the problem in the prior art where rigid gating of voice activity detection causes adaptive filters to be erroneously updated or frozen when the sound field changes, leading to cascaded distortion and sound quality degradation. It achieves smooth and continuous control of adaptive filtering and linkage adjustment of parameters between the preceding and following stages in call scenarios with various spatial characteristics, which can block cascaded degradation paths and stably output clean and clear voice signals.
[0005] This application provides an audio noise reduction method for a Bluetooth communication headset, including: acquiring a mixed audio stream containing human voice and background sound, and a reference audio stream mainly containing background sound, and separating and extracting audio amplitude data and temporarily stored audio phase data.
[0006] Based on audio amplitude data, spatial sound correlation coefficient, real-time sound energy deviation, and sound frequency distribution concentration are extracted for each discrete frequency point.
[0007] For each discrete frequency point, a frequency point update confidence value is generated to control the adjustment process;
[0008] The reference audio stream is used to perform background sound cancellation processing on the mixed audio stream. The confidence value of the frequency point is updated to control the dynamic adjustment range of the background sound cancellation processing at each frequency point. The error audio amplitude data after cancellation processing is output, and the overall working status index of the current processing time period is calculated simultaneously.
[0009] Based on the overall working status indicators, the historical data dependency weights in the sound clarity prediction process are dynamically adjusted;
[0010] Based on the adjusted historical data dependency weights, the error audio amplitude data is enhanced to obtain the target amplitude data, and then combined with the temporarily stored audio phase data for inverse transformation to obtain the time-series sound output signal.
[0011] Furthermore, the steps for extracting spatial sound correlation coefficients for each discrete frequency point include:
[0012] When processing audio data, for specific frequency nodes, the cross-correlation product data of the mixed audio stream and the reference audio stream is calculated, and stable correlation cross data is obtained through recursive smoothing along the time axis.
[0013] Simultaneously perform smoothing calculations on the audio amplitude data of the mixed audio stream and the reference audio stream to obtain smoothing energy values;
[0014] The absolute value of the correlation cross data in the stable state is divided by the square root of the product of the two smooth energy values, and the result of the division is used as the spatial sound correlation coefficient.
[0015] Further steps for determining real-time sound energy deviation and sound frequency distribution concentration include:
[0016] During periods of pure background noise in which no human voice activity was detected, the amplitude data of the mixed audio stream was smoothed to construct and slowly update long-term background sound energy prediction data.
[0017] The current amplitude data of the mixed audio stream is compared with the corresponding long-term estimated data by division to obtain the real-time sound energy deviation.
[0018] To measure the concentration of sound frequency distribution, the proportion of abnormal nodes whose amplitude exceeds the average baseline of adjacent frequency nodes in the mixed audio stream is statistically analyzed, and this proportion is used to measure the concentration of the overall sound frequency distribution.
[0019] The extracted feature parameters are summarized and output to the subsequent mapping and judgment stage for comprehensive cross-analysis.
[0020] Furthermore, the step of generating frequency point update confidence values for controlling the adjustment process includes:
[0021] A multi-dimensional feature evaluation logic is pre-built, and the spatial sound correlation coefficient and real-time sound energy deviation are introduced into the joint evaluation module for in-depth cross-validation;
[0022] When the correlation coefficient is high and the energy deviation is low, it is determined that the current sound is mainly background sound and the spatial characteristics are stable, and a high initial confidence value is generated as the offset control benchmark.
[0023] When the correlation coefficient is low or the energy deviation is high, it is determined that there is human voice mixed in or the spatial characteristics are seriously disordered, and a low initial confidence value is generated.
[0024] The initial confidence data is then completely output to the subsequent modulation stage for secondary dynamic correction calculations to ensure the reliability of the final data.
[0025] Furthermore, the steps for performing secondary dynamic correction calculations on the output in subsequent modulation stages to ensure the reliability of the final data include:
[0026] After obtaining the initial confidence data output by the joint evaluation module, the concentration of sound frequency distribution is introduced as a dynamic modulation reference factor to carry out secondary calibration.
[0027] When the concentration is in a high range, it indicates that the audio in the current time period is closer to the broadband background characteristics, and the initial confidence data corresponding to each frequency node should be appropriately increased.
[0028] When the concentration is in a low range, it indicates that there are human voice components with obvious harmonic structures in the audio, and the initial confidence data corresponding to each frequency node should be appropriately suppressed.
[0029] After the above modulation process, an intermediate adjustment value is obtained. This intermediate adjustment value is then subjected to boundary limiting constraint processing to keep it within a safe and continuous range, and is then issued as the final updated confidence value.
[0030] Furthermore, the step of performing background sound cancellation processing on the mixed audio stream using a reference audio stream includes:
[0031] The frequency domain data of the reference audio stream is used as the cancellation reference input, and the frequency domain data of the mixed audio stream is used as the desired response. The reference input is filtered by adaptive filtering logic to obtain the predicted data of the background sound.
[0032] The estimated data is subtracted from the mixed audio stream frequency domain data to obtain the error audio data. The rigid mechanism of uniform pace across the entire frequency band is abandoned during the iteration of the filtering logic operation parameters.
[0033] The parameter controlling the update amplitude is replaced with the frequency update confidence value corresponding to the current frequency node multiplied by the base step size value to obtain a dynamically adjustable value.
[0034] Furthermore, the steps for synchronously calculating the overall work status indicators for the current processing time period include:
[0035] After the filtering processing logic completes parameter iteration and updates and outputs error data, a macro-level status assessment and monitoring process is initiated to assess the adaptation quality and stability of the processing logic itself.
[0036] Collect the updated confidence values of all frequency nodes within the current processing time period, perform global weighted average calculation, and directly set the calculated results as the basic reference values of the status indicators.
[0037] Extract the updated confidence values of adjacent historical time periods for cross-period comparison, and statistically analyze the real-time distribution of the number of abnormal frequency nodes whose changes exceed the safety limit among all nodes.
[0038] If the percentage is too high, a downward correction will be applied to the base reference value of the status indicator. If the percentage is in a low and safe range, the status will be maintained or slightly adjusted, and the corrected value will be confirmed as the overall work status indicator.
[0039] Furthermore, the steps for dynamically adjusting the historical data-dependent weights in the sound intelligibility prediction process include:
[0040] The error audio data is sent to the sound clarity enhancement stage with a built-in guided prediction mechanism, and the historical dependency weights in the guided prediction mechanism are changed in real time according to the received overall working status indicators.
[0041] When the overall working status index is in a low range, clearly indicating that the external spatial environment has undergone complex changes or the front-end adaptation quality has declined significantly, the allocation share of historical data dependence weight should be proactively increased.
[0042] This operation makes the sharpness prediction process rely more heavily on historical baseline information corrected from the previous time period, thereby significantly reducing the reliance on real-time observed waveform data for the current time period.
[0043] Furthermore, the steps of inversely transforming the temporarily stored audio phase data to obtain the time-series audio output signal include:
[0044] Based on dynamically adjusted ambient sound amplitude and clarity prediction data, an advanced spectrum gain adjustment function is calculated for each independent frequency node in the current time period, which is precisely corresponding to it.
[0045] The advanced spectral gain adjustment function is multiplied with the error amplitude data output by the preamplifier at each frequency node to remove residual noise and extract the target amplitude data after deep purification.
[0046] The target amplitude data is recombined and matched with the audio phase data that was separated and temporarily stored in the initial conversion stage of the system to construct an enhanced frequency domain composite feature data stream with complete structure and accurate phase.
[0047] The enhanced frequency domain composite feature data stream is subjected to inverse time domain restoration transformation to obtain discrete time domain signal waveforms. Then, the signal waveforms in adjacent time periods are superimposed to eliminate boundary breaks, and finally, the denoised sound signal is output.
[0048] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:
[0049] By calculating the spatial correlation of the frequency division points of the mixed audio stream and the reference audio stream, and combining the real-time energy deviation and frequency distribution characteristics to generate continuously changing confidence values, this replaces the traditional rigid switching method based on fixed criteria. This avoids misjudging the current sound scene when the call sound field changes due to spatial characteristics such as being near a wall or inside a vehicle, ensuring that subsequent processing always obtains accurate and reliable control basis.
[0050] Furthermore, when performing background noise cancellation processing on the mixed audio stream using the reference audio stream, the confidence value is directly used to control the dynamic adjustment amplitude of the cancellation processing at each frequency point, thus achieving a smooth transition in the adjustment process. This prevents mislearning when there are human voices, which could lead to the partial elimination of human voice components. At the same time, it can track and adjust in a timely manner when the noise characteristics change, thereby maintaining a stable balance between protecting the integrity of human voices and eliminating background noise.
[0051] Furthermore, after the cancellation process is completed, the macroscopic working status index generated based on the confidence level value is used to dynamically adjust the dependence ratio between historical data and new observation data in the subsequent clarity enhancement stage. When the front-end processing status is detected to be unstable, the guiding role of historical information is automatically strengthened to suppress oscillation, thereby blocking the cascading path of front-end distortion to the back-end and avoiding abnormal noises such as water flow sounds and artificial whistles caused by processing mismatch in the final output signal. Attached Figure Description
[0052] Figure 1This is a flowchart illustrating an audio noise reduction method for a Bluetooth communication headset, as provided in an embodiment of this application. Detailed Implementation
[0053] This application provides an audio noise reduction method for Bluetooth communication headsets, solving the problem in the prior art where rigid gating of voice activity detection causes adaptive filters to be erroneously updated or frozen when the sound field changes, leading to cascaded distortion and sound quality degradation. By performing joint analysis of the spatial correlation degree and energy characteristics of the mixed audio stream and the reference audio stream on a frequency-by-frequency basis, a continuously updated confidence level is generated to smoothly adjust the frequency step size of the adaptive filter. Based on the filter state index, the historical dependency weight of the post-enhancement stage is dynamically adjusted, realizing smooth and continuous control of the adaptive filter and linkage adjustment of parameters of the preceding and following stages in call scenarios with various spatial characteristics changes. This fundamentally blocks the cascaded degradation path and stably outputs a clean and clear voice signal.
[0054] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0055] like Figure 1 The diagram shows a flowchart of an audio noise reduction method for a Bluetooth communication headset provided in an embodiment of this application. The method includes the following steps: acquiring a mixed audio stream containing human voice and background sound, and a reference audio stream mainly containing background sound; performing time-domain to frequency-domain conversion processing on the two audio streams respectively; and separating and extracting their respective audio amplitude data and temporarily stored audio phase data.
[0056] Based on the audio amplitude data of the two audio streams, the spatial sound correlation coefficient, real-time sound energy deviation, and sound frequency distribution concentration are extracted for each discrete frequency point.
[0057] For each discrete frequency point, the extracted spatial sound correlation coefficient, real-time sound energy deviation, and sound frequency distribution concentration are input into a preset mapping logic to generate a frequency point update confidence value for controlling the adjustment process.
[0058] The reference audio stream is used to perform background sound cancellation processing on the mixed audio stream. The confidence value of the frequency point is updated to control the dynamic adjustment range of the background sound cancellation processing at each frequency point. The error audio amplitude data after cancellation processing is output, and the overall working status index of the current processing time period is calculated simultaneously.
[0059] Based on the overall working status indicators, the historical data dependency weights in the sound clarity prediction process are dynamically adjusted;
[0060] Based on the adjusted historical data dependency weights, the error audio amplitude data is enhanced to obtain the target amplitude data, and then combined with the temporarily stored audio phase data for inverse transformation to obtain the time-series sound output signal.
[0061] In this embodiment, in practical applications, the headphones typically contain two microphones: a primary microphone and a secondary microphone. The primary microphone is located near the user's mouth and picks up a mixed audio stream containing the user's voice and ambient noise; the secondary microphone faces outward and primarily picks up ambient noise as a reference audio stream.
[0062] The system first segments the continuous sound waveform into extremely short time segments (e.g., 20 milliseconds per segment, with adjacent segments overlapping by 10 milliseconds to ensure continuity). Then, these time-domain waveforms are transformed to the frequency level, breaking down the complex sound into individual frequency nodes from low to high. During this process, amplitude data (representing the volume of each frequency) and phase data (representing the temporal order of the sound) are separated. The phase data is temporarily stored in the system's cache for later use.
[0063] Next, instead of simply and crudely cutting off noise, each frequency node is meticulously refined. Three key indicators are then extracted: spatial sound correlation coefficient, real-time sound energy deviation, and sound frequency distribution concentration. Through a pre-defined evaluation mechanism, these three indicators are converted into a "frequency update confidence value" (typically ranging from 0 to 1). This value acts like a smart accelerator, precisely controlling the operational amplitude of subsequent background noise cancellation stages. After cancellation processing, the preliminarily cleaned error audio amplitude data is output, and an overall operational status indicator representing the stability of the current processing is calculated.
[0064] Finally, the intensity of subsequent sound purification is dynamically adjusted based on this status indicator, and the processed amplitude data is reassembled with the previously stored phase data to restore a continuous sound waveform that the human ear can understand.
[0065] Furthermore, the steps for extracting spatial sound correlation coefficients for each discrete frequency point include:
[0066] When processing audio data, for specific frequency nodes, the cross-correlation product data of the mixed audio stream and the reference audio stream is calculated, and stable correlation cross data is obtained through recursive smoothing along the time axis.
[0067] Simultaneously perform smoothing calculations on the audio amplitude data of the mixed audio stream and the reference audio stream to obtain smoothing energy values;
[0068] The absolute value of the correlation cross data in the stable state is divided by the square root of the product of the two smooth energy values. The result of the division is used as the spatial sound correlation coefficient, which is used to characterize the degree of spatial correlation of the sound. The obtained parameter is then output to the subsequent logic.
[0069] This step can effectively avoid misjudging sound space in complex environments.
[0070] In this embodiment, to determine whether a sound at a certain frequency originates from the mouth or is transmitted from the outside, it is necessary to calculate the correlation between the two audio signals. The specific operation is as follows:
[0071] At specific frequency nodes, the mixed audio stream and the reference audio stream are cross-multiplied. To prevent drastic data jitter caused by momentary pops (such as the sound of wind), a smooth transition calculation method is used, the logic of which can be expressed by the following formula:
[0072] ;
[0073] in, This represents the current smoothed cross-correlation data. This indicates the cross-correlation data retained from the previous time period. This represents the product result calculated at the current instant. This is the smoothing adjustment coefficient.
[0074] Considering the processing power of the audio chip used, The value is typically set between 0.8 and 0.95. If the environment in which the headphones are used changes rapidly (such as while cycling), the [value] can be adjusted. Adjust the setting to a smaller value to speed up the response; if it is a quiet indoor environment, adjust it to a larger value for stability, leaving ample room for flexibility in actual product testing.
[0075] Subsequently, the smoothing values of the two audio energy streams are calculated simultaneously. Finally, the absolute value of the cross-correlation data is divided by the square root of the product of the two energy smoothing values, and the result is the "spatial sound correlation coefficient". The closer this value is to 1, the more likely the sound at that frequency is the same on both sides, which is likely environmental noise; the closer it is to 0, the greater the difference, which is likely human voice.
[0076] Further steps for determining real-time sound energy deviation and sound frequency distribution concentration include:
[0077] During periods of pure background noise in which no human voice activity was detected, the amplitude data of the mixed audio stream was smoothed to construct and slowly update long-term background sound energy prediction data.
[0078] The current amplitude data of the mixed audio stream is divided and compared with the corresponding long-term estimated data to obtain the real-time sound energy deviation, so as to objectively reflect the abnormal state of real-time energy deviating from the long-term stable background.
[0079] To measure the concentration of sound frequency distribution, the proportion of abnormal nodes whose amplitude exceeds the average baseline of adjacent frequency nodes in the mixed audio stream is statistically analyzed, and this proportion is used to measure the concentration of the overall sound frequency distribution.
[0080] The extracted feature parameters are summarized and output to the subsequent mapping and judgment stage for comprehensive cross-analysis.
[0081] In this embodiment, the methods for obtaining "real-time sound energy deviation" and "sound frequency distribution concentration" are mainly discussed.
[0082] First, there's the energy deviation: During a completely quiet period when the user isn't speaking, a long-term background sound energy estimate, or ambient noise level, is silently recorded and slowly updated. When processing the current sound segment, the volume of the current segment is divided by this long-term noise level. If the ratio is large, it indicates a sudden, noticeable sound, possibly a human voice or sudden noise; this is called energy deviation.
[0083] Secondly, there's the concentration of frequency distribution: human voices have a regular frequency distribution with high and low harmonics, while pure background noise is often chaotic and includes various frequencies. By examining each frequency node in the current sound segment, if the volume of a node exceeds the average of its left and right adjacent nodes, it's marked as a prominent node. The proportion of these prominent nodes in the total number of nodes is then calculated to determine whether the sound is concentrated like human voices or dispersed like background noise. These two characteristics provide multi-faceted evidence for subsequent judgments, avoiding the pitfalls of misjudgment that can occur with a single feature.
[0084] Furthermore, the step of generating frequency point update confidence values for controlling the adjustment process includes:
[0085] A multi-dimensional feature evaluation logic is pre-built, and the spatial sound correlation coefficient and real-time sound energy deviation are introduced into the joint evaluation module for in-depth cross-validation;
[0086] When the correlation coefficient is high and the energy deviation is low, it is determined that the current sound is mainly background sound and the spatial characteristics are stable, and a high initial confidence value is generated as the offset control benchmark.
[0087] When the correlation coefficient is low or the energy deviation is high, it is determined that there is voice mixed in or the spatial characteristics are seriously disordered, and a low initial confidence value is generated to avoid damage to the voice signal.
[0088] The initial confidence data is then completely output to the subsequent modulation stage for secondary dynamic correction calculations to ensure the reliability of the final data.
[0089] This embodiment discloses how to generate initial confidence data using extracted features. This is actually a multi-condition logical adjudication process.
[0090] A joint evaluation module is constructed. When a high spatial sound correlation coefficient is found at a certain frequency node, meaning the sounds heard by the microphones on both sides are very similar, and the real-time sound energy deviation is very low, meaning the current volume and background noise are similar, the evaluation module concludes that this part of the sound is pure and stable environmental noise. At this point, a relatively high initial confidence score is assigned (e.g., around 0.8).
[0091] Conversely, if the correlation coefficient is very low or the energy deviation is high, the judging module will assume that the user is speaking or that there has been a sudden and drastic change in ambient noise (such as suddenly opening a door into a noisy corridor). To protect human voices from being incorrectly eliminated, an extremely low initial confidence score is assigned (e.g., close to 0). This cross-validation mechanism reduces the risk of misjudgment caused by relying solely on volume.
[0092] Furthermore, the steps for performing secondary dynamic correction calculations on the output in subsequent modulation stages to ensure the reliability of the final data include:
[0093] After obtaining the initial confidence data output by the joint evaluation module, the concentration of sound frequency distribution is introduced as a dynamic modulation reference factor to carry out secondary calibration.
[0094] When the concentration is in a high range, it indicates that the audio in the current time period is closer to the broadband background characteristics, and the initial confidence data corresponding to each frequency node should be appropriately increased.
[0095] When the concentration is in a low range, it indicates that there are human voice components with obvious harmonic structures in the audio, and the initial confidence data corresponding to each frequency node should be appropriately suppressed.
[0096] After the above modulation process, an intermediate adjustment value is obtained. This intermediate adjustment value is then subjected to boundary limiting constraint processing to keep it within a safe and continuous range, and is then issued as the final updated confidence value.
[0097] In this embodiment, we further explain how to perform secondary refinement on the initial confidence data.
[0098] After receiving the initial judgment of the output, a third feature, the concentration of sound frequency distribution, is introduced for verification. If the concentration value indicates that the overall sound characteristics are very chaotic (like wind or traffic noise), the overall environment is considered to be harsh, and an instruction is issued to slightly increase the initial confidence level to enhance noise suppression. If the concentration value indicates that the sound is very regular (clearly possessing the timbre characteristics of human voices), the initial confidence level is forcibly lowered to protect sound quality.
[0099] After this secondary calibration, an intermediate adjustment value is obtained. To prevent this value from deviating and causing system crashes, a boundary limiting constraint is added to forcibly restrict this value to a safe, continuous range (e.g., between 0.01 and 0.99). The value output by this logic is the final frequency point update confidence value that is sent out for execution.
[0100] Furthermore, the step of performing background sound cancellation processing on the mixed audio stream using a reference audio stream includes:
[0101] The frequency domain data of the reference audio stream is used as the cancellation reference input, and the frequency domain data of the mixed audio stream is used as the desired response. The reference input is filtered by adaptive filtering logic to obtain the predicted data of the background sound.
[0102] The estimated data is subtracted from the mixed audio stream frequency domain data to obtain the error audio data. The rigid mechanism of uniform pace across the entire frequency band is abandoned during the iteration of the filtering logic operation parameters.
[0103] The parameter controlling the update amplitude is replaced by multiplying the frequency update confidence value corresponding to the current frequency node with the base step size value to obtain a dynamically adjustable value, thereby realizing continuous adaptive control of a single node.
[0104] A higher confidence level drives a larger parameter update, while a lower confidence level suppresses the update or brings it close to a stop state. In this way, error audio data with residual features is output for subsequent processing.
[0105] Furthermore, the steps for synchronously calculating the overall work status indicators for the current processing time period include:
[0106] After the filtering processing logic completes parameter iteration and updates and outputs error data, a macro-level status assessment and monitoring process is initiated to assess the adaptation quality and stability of the processing logic itself.
[0107] Collect the updated confidence values of all frequency nodes within the current processing time period, perform global weighted average calculation, and directly set the calculated results as the basic reference values of the status indicators.
[0108] Extract the updated confidence values of adjacent historical time periods for cross-period comparison, and statistically analyze the real-time distribution of the number of abnormal frequency nodes whose changes exceed the safety limit among all nodes.
[0109] If the percentage is too high, a downward correction will be applied to the base reference value of the status indicator. If the percentage is in a low and safe range, the status will be maintained or slightly adjusted, and the corrected value will be confirmed as the overall work status indicator.
[0110] In this embodiment, after the front-end filter completes one round of work, it is necessary to evaluate its performance. First, the updated confidence values of all frequency nodes within the current time period are collected, an average score is calculated, and this average score is used as the basic reference value.
[0111] Next, review the historical data and compare it with data from previous time periods to see if any frequency confidence values have experienced abrupt changes. If a large number of frequency nodes show extremely large jumps, it's determined that the current soundstage environment is highly unstable, and the front-end filter may be unable to keep up. Therefore, the baseline reference score is deducted, applying a downward correction. Conversely, if everything is stable, the score remains unchanged. The final score is the overall performance indicator, objectively reflecting the headphones' ability to handle complex environments.
[0112] Furthermore, the steps for dynamically adjusting the historical data-dependent weights in the sound intelligibility prediction process include:
[0113] The error audio data is sent to the sound clarity enhancement stage with a built-in guided prediction mechanism, and the historical dependency weights in the guided prediction mechanism are changed in real time according to the received overall working status indicators.
[0114] When the overall working status index is in a low range, clearly indicating that the external spatial environment has undergone complex changes or the front-end adaptation quality has declined significantly, the allocation share of historical data dependence weight should be proactively increased.
[0115] This operation makes the sharpness prediction process rely more heavily on historical baseline information corrected from the previous time period, thereby significantly reducing the reliance on real-time observed waveform data for the current time period.
[0116] Through the aforementioned defensive adaptive adjustment strategy, the predicted results are effectively avoided from drastic distortion fluctuations caused by temporary mismatch in front-end processing, ensuring a natural and smooth sound experience even in complex sound fields.
[0117] In this embodiment, although the audio data processed previously has most of the noise removed, some subtle noise often remains, requiring a subsequent sound clarity enhancement stage, typically using noise reduction and enhancement algorithms. A key setting in this stage is called historical data dependency weighting.
[0118] When the overall work status indicators from the previous physical examination are very low, it indicates that the environment is chaotic and the front-end processing is struggling, so a defense mechanism is activated. The weighting of historical data is proactively increased. By increasing reliance on historical patterns, the algorithm effectively avoids producing robotic, metallic, or bubbly sounds due to sudden garbled text interference. Even in extremely noisy environments, residual noise is smoothed out, ensuring that human voices sound natural and even.
[0119] Furthermore, the steps of inversely transforming the temporarily stored audio phase data to obtain the time-series audio output signal include:
[0120] Based on dynamically adjusted ambient sound amplitude and clarity prediction data, an advanced spectrum gain adjustment function is calculated for each independent frequency node in the current time period, which is precisely corresponding to it.
[0121] The advanced spectral gain adjustment function is multiplied with the error amplitude data output by the preamplifier at each frequency node to remove residual noise and extract the target amplitude data after deep purification.
[0122] The target amplitude data is recombined and matched with the audio phase data that was separated and temporarily stored in the initial conversion stage of the system to construct an enhanced frequency domain composite feature data stream with complete structure and accurate phase.
[0123] The enhanced frequency domain composite feature data stream is subjected to inverse time domain restoration transformation to obtain discrete time domain signal waveforms. Then, the signal waveforms in adjacent time periods are superimposed to eliminate boundary breaks, and finally, the denoised sound signal is output.
[0124] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0125] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.
[0126] Those skilled in the art will recognize that the modules and algorithm steps of the various examples 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 implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0127] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.
[0128] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0129] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An audio noise reduction method for a Bluetooth communication headset, characterized in that, Includes the following steps: Acquire a mixed audio stream containing human voice and background sound, as well as a reference audio stream dominated by background sound, and separate and extract the audio amplitude data and the temporarily stored audio phase data. Based on audio amplitude data, spatial sound correlation coefficient, real-time sound energy deviation, and sound frequency distribution concentration are extracted for each discrete frequency point. For each discrete frequency point, a frequency point update confidence value is generated to control the adjustment process; The reference audio stream is used to perform background sound cancellation processing on the mixed audio stream. The confidence value of the frequency point is updated to control the dynamic adjustment range of the background sound cancellation processing at each frequency point. The error audio amplitude data after cancellation processing is output, and the overall working status index of the current processing time period is calculated simultaneously. Based on the overall working status indicators, the historical data dependency weights in the sound clarity prediction process are dynamically adjusted; Based on the adjusted historical data dependency weights, the error audio amplitude data is enhanced to obtain the target amplitude data, and then combined with the temporarily stored audio phase data for inverse transformation to obtain the time-series sound output signal.
2. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps for extracting spatial sound correlation coefficients for each discrete frequency point include: When processing audio data, for specific frequency nodes, the cross-correlation product data of the mixed audio stream and the reference audio stream is calculated, and stable correlation cross data is obtained through recursive smoothing along the time axis. Simultaneously perform smoothing calculations on the audio amplitude data of the mixed audio stream and the reference audio stream to obtain smoothing energy values; The absolute value of the correlation cross data in the stable state is divided by the square root of the product of the two smooth energy values, and the result of the division is used as the spatial sound correlation coefficient.
3. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps for real-time sound energy deviation and sound frequency distribution concentration include: During periods of pure background noise in which no human voice activity was detected, the amplitude data of the mixed audio stream was smoothed to construct and slowly update long-term background sound energy prediction data. The current amplitude data of the mixed audio stream is compared with the corresponding long-term estimated data by division to obtain the real-time sound energy deviation. To measure the concentration of sound frequency distribution, the proportion of abnormal nodes whose amplitude exceeds the average baseline of adjacent frequency nodes in the mixed audio stream is statistically analyzed, and this proportion is used to measure the concentration of the overall sound frequency distribution. The extracted feature parameters are summarized and output to the subsequent mapping and judgment stage for comprehensive cross-analysis.
4. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps for generating frequency point update confidence values for controlling the regulation process include: A multi-dimensional feature evaluation logic is pre-built, and the spatial sound correlation coefficient and real-time sound energy deviation are introduced into the joint evaluation module for in-depth cross-validation; When the correlation coefficient is high and the energy deviation is low, it is determined that the current sound is mainly background sound and the spatial characteristics are stable, and a high initial confidence value is generated as the offset control benchmark. When the correlation coefficient is low or the energy deviation is high, it is determined that there is human voice mixed in or the spatial characteristics are seriously disordered, and a low initial confidence value is generated. The initial confidence data is then completely output to the subsequent modulation stage for secondary dynamic correction calculations to ensure the reliability of the final data.
5. The audio noise reduction method for a Bluetooth communication headset as described in claim 4, characterized in that, The steps for performing secondary dynamic correction calculations on the output in subsequent modulation stages to ensure the reliability of the final data include: After obtaining the initial confidence data output by the joint evaluation module, the concentration of sound frequency distribution is introduced as a dynamic modulation reference factor to carry out secondary calibration. When the concentration is in a high range, it indicates that the audio in the current time period is closer to the broadband background characteristics, and the initial confidence data corresponding to each frequency node should be appropriately increased. When the concentration is in a low range, it indicates that there are human voice components with obvious harmonic structures in the audio, and the initial confidence data corresponding to each frequency node should be appropriately suppressed. After the above modulation process, an intermediate adjustment value is obtained. This intermediate adjustment value is then subjected to boundary limiting constraint processing to keep it within a safe and continuous range, and is then issued as the final updated confidence value.
6. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps of performing background noise cancellation processing on the mixed audio stream using a reference audio stream include: The frequency domain data of the reference audio stream is used as the cancellation reference input, and the frequency domain data of the mixed audio stream is used as the desired response. The reference input is filtered by adaptive filtering logic to obtain the predicted data of the background sound. The estimated data is subtracted from the mixed audio stream frequency domain data to obtain the error audio data. The rigid mechanism of uniform pace across the entire frequency band is abandoned during the iteration of the filtering logic operation parameters. The parameter controlling the update amplitude is replaced with the frequency update confidence value corresponding to the current frequency node multiplied by the base step size value to obtain a dynamically adjustable value.
7. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps for synchronously calculating the overall work status indicators for the current processing time period include: After the filtering processing logic completes parameter iteration and updates and outputs error data, a macro-level status assessment and monitoring process is initiated to assess the adaptation quality and stability of the processing logic itself. Collect the updated confidence values of all frequency nodes within the current processing time period, perform global weighted average calculation, and directly set the calculated results as the basic reference values of the status indicators. Extract the updated confidence values of adjacent historical time periods for cross-period comparison, and statistically analyze the real-time distribution of the number of abnormal frequency nodes whose changes exceed the safety limit among all nodes. If the percentage is too high, a downward correction will be applied to the base reference value of the status indicator. If the percentage is in a low and safe range, the status will be maintained or slightly adjusted, and the corrected value will be confirmed as the overall work status indicator.
8. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps involved in dynamically adjusting the historical data-dependent weights in the sound clarity prediction process include: The error audio data is sent to the sound clarity enhancement stage with a built-in guided prediction mechanism, and the historical dependency weights in the guided prediction mechanism are changed in real time according to the received overall working status indicators. When the overall working status index is in a low range, clearly indicating that the external spatial environment has undergone complex changes or the front-end adaptation quality has declined significantly, the allocation share of historical data dependence weight should be proactively increased. This operation makes the sharpness prediction process rely more heavily on historical baseline information corrected from the previous time period, thereby significantly reducing the reliance on real-time observed waveform data for the current time period.
9. The audio noise reduction method for a Bluetooth communication headset as described in claim 1, characterized in that, The steps for obtaining a time-series audio output signal by inversely transforming the temporarily stored audio phase data include: Based on dynamically adjusted ambient sound amplitude and clarity prediction data, an advanced spectrum gain adjustment function is calculated for each independent frequency node in the current time period, which is precisely corresponding to it. The advanced spectral gain adjustment function is multiplied with the error amplitude data output by the preamplifier at each frequency node to remove residual noise and extract the target amplitude data after deep purification. The target amplitude data is recombined and matched with the audio phase data that was separated and temporarily stored in the initial conversion stage of the system to construct an enhanced frequency domain composite feature data stream with complete structure and accurate phase. The enhanced frequency domain composite feature data stream is subjected to inverse time domain restoration transformation to obtain discrete time domain signal waveforms. Then, the signal waveforms in adjacent time periods are superimposed to eliminate boundary breaks, and finally, the denoised sound signal is output.