Big data-based bluetooth earphone use situation analysis method and system

Abstract

This invention discloses a method and system for analyzing Bluetooth headset usage based on big data, relating to the field of signal processing technology. The method involves acquiring and processing a time-domain observation vector; performing a Fast Fourier Transform (FFT) on the time-domain observation vector to obtain a frequency-domain amplitude sequence and the original phase; estimating the observation noise variance of the frequency-domain amplitude sequence using the minimum value tracking method; calculating the objective acoustic indices of the current frame; and processing the objective acoustic indices through a hierarchical decision-making mechanism to obtain risk-sensitive parameters; extracting the risk-sensitive parameters, the frequency-domain amplitude sequence, and the original phase; and calculating a first estimated amplitude spectrum using CRSFIR frequency-domain amplitude state estimation; finally, performing voice activity detection and classification processing based on the first estimated amplitude spectrum to obtain the final usage results. This invention effectively improves the noise suppression capability of Bluetooth headsets, adaptively updates risk environment parameters in real time, improves the Bluetooth call experience, and reduces headset power consumption management.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method and system for analyzing Bluetooth headset usage based on big data. Background Technology

[0002] With the rapid development of wireless communication technology and the widespread use of smart mobile terminals, Bluetooth headsets have become an indispensable audio device in people's daily lives. However, in actual use, Bluetooth headsets often face complex and varied acoustic environment challenges such as traffic noise, human voice interference, wind noise, and indoor reverberation. These noises seriously reduce the clarity and intelligibility of voice communication and affect the user experience.

[0003] Currently, audio enhancement technologies for Bluetooth headsets mainly focus on noise suppression, echo cancellation, and speech enhancement. Traditional methods include spectral subtraction, Wiener filtering, statistical model-based methods, and, more recently, deep learning methods. Spectral subtraction estimates the noise spectrum and subtracts it from the noisy speech spectrum, which is simple but prone to generating musical noise and causing speech distortion. Wiener filtering is optimized under the minimum mean square error criterion, but its adaptability to noise non-stationarity is limited. Hidden Markov Model (HMM) methods based on statistical models require accurate prior knowledge and have insufficient generalization ability in practical applications. Although deep learning methods have shown great potential in noise suppression, their high computational complexity and model size limit their deployment and application on resource-constrained Bluetooth headset embedded platforms. In terms of filtering theory, traditional Kalman filters belong to the infinite impulse response structure, and their performance is highly dependent on accurate system models and noise statistics. They are not robust to model mismatch and sudden noise, and are prone to error accumulation or even filter divergence. Finite impulse response (FIR) filters exhibit stronger robustness by using only observation data within the most recent finite time window, thus avoiding the accumulation of historical errors. However, designing appropriate cost functions and optimization criteria to achieve a balance between performance and robustness in complex noise environments remains a challenge. Furthermore, the inherent acoustic echo of the device needs to be effectively eliminated. Secondly, residual noise in the ear canal after active noise cancellation needs to be further suppressed. Finally, the device needs to adaptively adjust its processing strategy in different scenarios. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies have difficulty in reducing noise from non-stationary sudden noise, have poor robustness, and traditional filters cannot adapt to complex and changing acoustic environments. Furthermore, their performance is heavily dependent on accurate noise models and parameter adjustments are difficult. Deep learning-based models or complex optimization algorithms have high computational complexity, making it difficult to accurately control the real-time noise reduction of Bluetooth headsets.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: The first aspect is a method for analyzing Bluetooth headset usage based on big data, which includes the following steps: Step S1: Acquire the original multi-channel speech signal and reference signal and perform preprocessing to obtain the time-domain observation vector; Step S2: Perform a fast Fourier transform on the time-domain observation vector to obtain the corresponding frequency-domain observation vector and extract the amplitude of each frequency point to form a frequency-domain amplitude sequence and the original phase; Step S3: The minimum tracking method is used to estimate the observation noise variance of the frequency domain amplitude sequence, the objective acoustic index of the current frame is calculated, and the objective acoustic index is processed through a hierarchical decision-making mechanism to obtain risk-sensitive parameters. Step S4: Extract risk-sensitive parameters, frequency domain amplitude sequence and original phase, and calculate the first estimated amplitude spectrum through frequency domain amplitude state estimation processing of CRSFIR; Step S5: Based on the first estimated amplitude spectrum, perform speech activity detection and classification to obtain the final usage result.

[0006] Preferably, step S1 includes the following sub-steps: Step S11: The original multi-channel voice signal is synchronously acquired through the built-in microphone array of the Bluetooth headset, and a reference signal is acquired through the speaker. The original multi-channel voice signal includes environmental noise and residual noise. Step S12: The multi-channel speech signal and the reference signal are digitized by an analog-to-digital converter to obtain a discrete time series. Step S13: Use an adaptive filter to perform acoustic echo cancellation on the digitized reference signal to obtain a multi-channel de-echo signal; Step S14: Based on the discrete time series, spatial filtering is performed on the multi-channel de-echo signal using a beamforming algorithm to obtain a single-channel enhanced signal; Step S15: Overlap the single-channel enhancement signal into frames according to a fixed frame length and frame shift to obtain the enhancement signal of each frame. Multiply the enhancement signal point by point using a window function vector of the same frame length to obtain a time-domain observation vector. Arrange each time-domain observation vector in the order of each frame to obtain a time-domain observation matrix (L*K), where L is the row of the time-domain observation matrix and K is the column of the time-domain observation matrix. The ambient noise is the noise of the current environment that is about to be transmitted into the ear, collected by the reference microphone; The residual noise refers to the residual noise near the inner ear canal of the earphone after noise reduction processing.

[0007] Preferably, step S2 includes the following sub-steps: Step S21: Perform a Fast Fourier Transform on the time-domain observation vector of each frame to obtain the frequency-domain observation vector. Step S22: Extract the amplitude for each frequency point in the frequency domain observation vector to obtain the frequency domain amplitude sequence, and extract the phase of each frequency point as the original phase and save it to the phase buffer.

[0008] Preferably, step S3 includes the following sub-steps: Step S31: The minimum value tracking method is used to calculate the observation noise of the frequency domain amplitude sequence to obtain the observation noise variance; Step S32: Calculate the objective acoustic metrics for the current frame. The objective acoustic metrics include full-band signal-to-noise ratio, frequency band signal-to-noise ratio, noise stationarity measure, noise burst ratio, and estimation error statistics. Step S33: The objective acoustic indicators are dynamically calculated through a hierarchical decision-making mechanism to obtain the sensitive parameters of the current frame; Step S34: Perform first-order low-pass filtering smoothing on the sensitive parameters of the current frame and output the risk sensitive parameters.

[0009] Preferably, step S4 includes the following sub-steps: Step S41: Extract the frequency domain amplitude sequence of the most recent N frames at the current frequency point from the phase buffer and use it as the observation vector; Step S42: Define a state scalar for the observation vector corresponding to each frequency point, and establish a state space model to obtain state space data; Step S43: Input the state space data into the CRSFIR filter to calculate and obtain the instantaneous state estimation sequence; Step S44: Construct an extended matrix based on the state-space parameters, risk-sensitive parameters, and observation sequence; construct a system of linear equations based on the extended matrix and instantaneous state estimates; and calculate the amplitude estimate at the current frequency point. Step S45: Repeat steps S42, S43 and S44 for all frequency points, arrange the obtained amplitude estimates of each frequency point according to frequency, and combine them into the first estimated amplitude spectrum.

[0010] Preferably, step S42 specifically includes: By iterating through each frequency point, a state-space model is established using a first-order Markov transform based on the observation vector obtained in step S41. The computational expression of the state-space model is as follows: ; ; in For the first Time and Frequency The state scalar These are the state transition coefficients. The process noise is Gaussian distributed. For the first Time and Frequency The observation scalar, H is the observation coefficient (usually set to 1). The observation noise is Gaussian distributed. Based on the state equation of the state scalar, the initial state estimate is calculated using the least squares estimation method. The initial state estimate and the observation sequence are then combined. The parameters corresponding to the state space model are input into step S43 as state space data. The observation sequence is used to represent the N latest state estimation observations from the m-th frame to the k-th frame.

[0011] Preferably, step S43 specifically includes: Set the accumulation interval to a reasonable range within the observation window: Preset accumulation interval length The total window length is 70% of the total window length, which is the length of N frames. The start time s and end time k-1 of the accumulation interval are set. When the start time s is greater than or equal to m, the accumulation interval is within the observation window and the current accumulation interval is a valid interval. When the start time s is less than m, the accumulation interval is not within the observation window and the accumulation interval is converted into an interval with m as the start time and k-1 as the end time. For each time interval l, according to the cumulative interval length, the sub-observation vector and the corresponding observation noise variance vector are extracted from the observation sequence. The gain vector is then calculated using an instantaneous risk-sensitive filter. The expression for the gain vector is as follows: ; in, This represents the gain vector of the instantaneous risk-sensitive filter for the current accumulation interval length. These are the state transition coefficients. As a risk-sensitive parameter, This represents the average value of the observed noise variance vector; The instantaneous state estimate is calculated based on the gain vector, and the expression for calculating the instantaneous state estimate is as follows: ; in, For sub-observation vectors, This represents the gain vector of the instantaneous risk-sensitive filter for the current accumulation interval length. This is an instantaneous state estimate; The calculated instantaneous state estimates are stored in array form according to the time series to obtain the instantaneous state estimate sequence.

[0012] Preferably, step S44 specifically includes: Extract the instantaneous state estimation sequence, and construct the extended matrix corresponding to the current frequency point f using risk-sensitive parameters, sub-observation vectors, and parameters of the state-space model. The extended matrix has dimension [missing information]. The symmetric real matrix, the method for constructing the extended matrix is ​​as follows: The extended matrix is ​​divided into sub-matrix blocks, which include a top-left sub-matrix block, a middle sub-matrix block, and a bottom-right sub-matrix block. The top-left sub-matrix block represents the system state equation and the constraint relationship between process noise, observation noise, and risk sensitivity. The middle and bottom-right sub-matrix blocks reflect the coupling relationship between the instantaneous state estimation sequence, the current observation vector, and the risk sensitivity parameter. The right-hand vector is obtained by concatenating the instantaneous state estimation sequence and the observation vector at the current moment. The row indices of the extended matrix are divided into a first index set and a second index set. The first index set is the set of the global state and the current magnitude estimate. The second index set is the historical intermediate variable index set. The right-side vector is divided into a first right-side sub-vector and a second right-side sub-vector according to the first index set and the second index set. The first right-side sub-vector is the relevant component corresponding to the first index set, and the second right-side sub-vector is the relevant component corresponding to the second index set. Submatrices are extracted based on a first index set and a second index set. The submatrices include a first submatrice, a second submatrice, a third submatrice, and a fourth submatrice. The first submatrice is used to extract elements whose row index and column index are both in the first index set. The second submatrice is used to extract elements whose row index is in the first index set and whose column index is in the second index set. The third submatrice is the transpose of the second index. The fourth submatrice is used to extract elements whose row index and column index are both in the second index set. Based on the corresponding blocks of the right-hand vector, the reduced coefficient matrix and the low-dimensional linear equation system are constructed using the Schur complement form. The computational expressions for the reduced coefficient matrix and the low-dimensional linear equation system are as follows: ; ; ; in, This is the first submatrix. For the second submatrix, It is the inverse of the fourth submatrix. For the third submatrix, The first right sub-vector, The second right sub-vector, Let A be the parameter vector to be solved, and let A be a real symmetric matrix. The result vector; The amplitude estimate of the current frequency point f at time k is obtained by solving the low-dimensional linear equation system.

[0013] Preferably, step S5 includes the following sub-steps: Step S51: Perform speech activity detection on the first estimated amplitude spectrum to obtain speech detection results. The speech detection results include speech frames and non-speech frames. The speech activity detection includes calculating the frequency band energy of the first estimated amplitude spectrum for each frame, dividing the frequency band energy according to a preset speech threshold, marking the frame corresponding to the frequency band energy as a speech frame when the frequency band energy of the existing frame is greater than the preset speech threshold, and marking the frame corresponding to the frequency band energy as a non-speech frame when the frequency band energy of the existing frame is less than the preset speech threshold. Smooth the speech detection results to obtain a speech tag sequence. Step S52: Set a time window, and calculate the user temporal features based on the normalization of the speech tag sequence. The user temporal features include speech activity, average speech segment length, average silence segment length and number of speech segments, mean energy change rate and variance of energy change rate. The speech activity is the ratio of the number of speech frames in the current time window to the total number of speech frames. The processing logic for the average speech segment length is to extract a continuous speech marker sequence of speech frames as a speech segment, and calculate the average speech segment length by averaging the length of each speech segment. The processing logic for the average silence segment length is to extract a continuous speech marker sequence of non-speech frames as a non-speech segment, calculate the average length of all non-speech segments to obtain the average silence segment length, and count the number of speech segments in the current time window to obtain the number of speech segments. For a frame marked as a speech frame, the energy change rate between adjacent speech frames is calculated using the first estimated amplitude spectrum to obtain the mean and variance of the energy change rate. Step S53: Based on the user's temporal characteristics and the mean and variance of the energy change rate, pattern recognition is performed to obtain the final usage results. The final usage results include continuous dialogue communication, silent listening, and discrete interaction. A first threshold, a second threshold, and a third threshold are set. The first threshold is used to determine speech activity, the second threshold is used to determine the average silent segment length, and the third threshold is used to determine the average speech segment length. When the voice activity level is greater than or equal to the first threshold, the average silence segment length is less than the second threshold, and the average voice segment length is greater than or equal to the third threshold, the current time window is determined to be a continuous dialogue communication situation. When the voice activity level is less than the first threshold and the average length of the silent segment is greater than or equal to the second threshold, the current time window is determined to be a silent listening situation. The remaining time window that does not meet the conditions of continuous dialogue communication and silent listening is determined to be a discrete interaction situation.

[0014] Secondly, a Bluetooth headset usage analysis system based on big data includes a data acquisition module, a frequency domain analysis module, a risk-sensitive parameter estimation module, an amplitude estimation module, and a usage analysis module. The acquisition module is used to acquire multi-channel speech signals containing environmental noise and residual noise through the built-in microphone array of the earphone, and to acquire the speaker reference signal. The multi-channel speech signal and the reference signal are converted from analog to digital to obtain a discrete time series. The reference signal is acoustically echo-cancelled using an adaptive filter to obtain a multi-channel de-echo signal. The de-echo signal is beamformed to obtain a single-channel enhanced signal, and the signal is framed and windowed to construct a time-domain observation matrix. The frequency domain analysis module is used to perform a fast Fourier transform on the time domain observation vector of each frame to obtain the frequency domain observation vector, extract the amplitude based on each frequency point to form a frequency domain amplitude sequence, and extract the corresponding phase information; and save the original phase of each frequency point to the phase buffer. The risk-sensitive parameter estimation module is used to estimate the observation noise variance of the frequency domain amplitude sequence using the minimum value tracking method, calculate the objective acoustic indicators of the current frame, including full-band signal-to-noise ratio, frequency band signal-to-noise ratio, noise stationarity measure, noise burst ratio, and estimation error statistics. The objective acoustic indicators are mapped to the sensitive parameters of the current frame through a hierarchical decision mechanism, and the sensitive parameters are smoothed by first-order low-pass filtering to output the risk-sensitive parameters. The amplitude estimation module is used to extract risk-sensitive parameters, frequency domain amplitude sequences, and original phases, and calculate the first estimated amplitude spectrum through frequency domain amplitude state estimation processing of CRSFIR. The usage analysis module is used to detect and classify speech activity based on the first estimated amplitude spectrum to obtain the final usage results.

[0015] The beneficial effects of this invention: This invention mainly focuses on improving the ability of Bluetooth headsets to analyze user microphone-received voice in voice call scenarios. Through innovative signal processing and pattern recognition technologies, it significantly improves the clarity of voice calls in noisy environments, directly improving the core call experience. Traditional headset microphones often capture voice that is severely contaminated by environmental noise, making it difficult for the other party to hear. The core of this invention lies in employing cumulative risk-sensitive FIR (CRSFIR) filtering technology. This technology, through a dynamically adaptive risk-sensitive parameter θ, can intelligently distinguish between stable background noise and sudden interference, and applies stronger suppression to sudden noise. Combined with beamforming in step S1 to focus the direction of the human voice, the resulting first estimated amplitude spectrum greatly improves the purity of the user's speech. Even in complex noisy environments, users do not need to deliberately increase their volume to ensure clear speech. One's own voice is clearly picked up and transmitted to the other party, solving the problem of unclear hearing in Bluetooth voice calls. In addition, this invention is not simply headphone noise reduction, but intelligently distinguishes three key call-related modes through in-depth analysis of the enhanced voice: continuous dialogue communication, silent listening, and discrete interaction. This achieves accurate identification of the user's call behavior patterns, providing key decision-making basis for real-time call quality optimization and power consumption control. Microphone voice analysis must be real-time and low-latency. Although this invention adopts the advanced CRSFIR algorithm, it simplifies complex mathematical operations into fixed-dimensional calculations that can be efficiently executed on embedded chips by utilizing the Shure complement dimensionality reduction method with a special matrix structure. This allows for real-time analysis of call voice within the limited computing power and low power consumption budget of the headphones, ensuring the best call experience while reducing power consumption management of Bluetooth headphones. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a Bluetooth headset usage analysis method based on big data, provided as an embodiment of the present invention; Figure 2 This is a basic flowchart of a Bluetooth headset usage analysis system based on big data, provided as an embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1 It provides a method for analyzing Bluetooth headset usage based on big data, including the following steps: Step S1: Acquire the original multi-channel speech signal and reference signal and perform preprocessing to obtain the time-domain observation vector; Step S2: Perform a fast Fourier transform on the time-domain observation vector to obtain the corresponding frequency-domain observation vector and extract the amplitude of each frequency point to form a frequency-domain amplitude sequence and the original phase; Step S3: The minimum tracking method is used to estimate the observation noise variance of the frequency domain amplitude sequence, the objective acoustic index of the current frame is calculated, and the objective acoustic index is processed through a hierarchical decision-making mechanism to obtain risk-sensitive parameters. Step S4: Extract risk-sensitive parameters, frequency domain amplitude sequence and original phase, and calculate the first estimated amplitude spectrum through frequency domain amplitude state estimation processing of CRSFIR; Step S5: Based on the first estimated amplitude spectrum, perform speech activity detection and classification to obtain the final usage result.

[0019] To enhance the ability of Bluetooth headsets to analyze user microphone-received voice during voice calls, this invention employs innovative signal processing and pattern recognition technologies to significantly improve the clarity of voice calls in noisy environments, directly improving the core call experience. Traditional headset microphones often suffer from severe environmental noise contamination, making it difficult for the other party to hear clearly. The core of this invention lies in its use of Cumulative Risk-Sensitive First-Reference Filtering (CRSFIR). This technology, through a dynamically adaptive risk-sensitive parameter θ, intelligently distinguishes between stable background noise and sudden interference, and applies stronger suppression to sudden noise. Combined with beamforming in step S1 to focus on the direction of the user's voice, the resulting first estimated amplitude spectrum greatly improves the clarity of the user's speech. Even in complex noisy environments, users do not need to deliberately increase their volume to ensure their voice is clearly heard. Clearly picking up and transmitting audio to the other party solves the problem of unclear hearing during Bluetooth voice calls. Furthermore, this invention is not simply about noise reduction in headphones; it intelligently distinguishes three key call-related modes—continuous conversation, silent listening, and discrete interaction—through in-depth analysis of the enhanced audio. This enables accurate identification of user call behavior patterns, providing crucial decision-making support for real-time call quality optimization and power consumption control. Since microphone-based audio analysis must be real-time and low-latency, this invention, while employing the advanced CRSFIR algorithm, simplifies complex mathematical operations into fixed-dimensional calculations that can be efficiently executed on embedded chips by utilizing the Shure complement dimensionality reduction method with a special matrix structure. This allows for real-time analysis of call audio even within the limited computing power and low power consumption budget of the headphones, ensuring the best call experience while reducing Bluetooth headset power consumption.

[0020] Step S1 includes the following sub-steps: Step S11: The original multi-channel voice signal is synchronously acquired through the built-in microphone array of the Bluetooth headset, and a reference signal is acquired through the speaker. The original multi-channel voice signal includes ambient noise and residual noise. Step S12: The multi-channel speech signal and the reference signal are digitized by an analog-to-digital converter to obtain a discrete time series. Step S13: Use an adaptive filter to perform acoustic echo cancellation on the digitized reference signal to filter out echo interference generated by the speaker playback and obtain a multi-channel de-echo signal. Step S14: Based on the discrete time series, spatial filtering is performed on the multi-channel de-echo signal using a beamforming algorithm to obtain a single-channel enhanced signal; Step S15: Overlap the single-channel enhancement signal into frames according to a fixed frame length and frame shift to obtain the enhancement signal of each frame. Multiply the enhancement signal point by point using a window function vector of the same frame length to obtain a time-domain observation vector. Arrange each time-domain observation vector in the order of each frame to obtain a time-domain observation matrix (L*K), where L is the row of the time-domain observation matrix, used to represent the length of the fixed frame length, and K is the column of the time-domain observation matrix, used to represent the total number of frames of the enhancement signal. The ambient noise is the noise of the current environment that is about to be transmitted into the ear, as collected by the reference microphone; The residual noise is the residual noise near the inner ear canal of the earphone after noise reduction processing.

[0021] In this embodiment, multiple signals, including ambient noise, residual noise, and speaker reference signals, are simultaneously acquired using a spatially distributed microphone array. Then, the adaptive filter coefficients are updated in real time using MLMS or RLS algorithms. The speaker signals along the acoustic path are subtracted from the microphone signals to obtain multi-channel echo-reducing signals. The weighted sum of each microphone signal is calculated using a minimum variance distortion-free response algorithm to enhance the signal in the target direction and suppress interference from other directions. Finally, a fixed frame length of 25ms and a frame shift of 10ms are set using a window function to convert the continuous signal into a time-frequency observation matrix that facilitates time-frequency analysis. Information about the spatial sound field is acquired through multi-channel acquisition, providing a data foundation for beamforming. For the echo cancellation in step S13, the echo is prevented from being mistakenly identified as the target speech signal. The single-channel enhanced signal is segmented according to the frame length using a Hamming window function to obtain a time-domain observation vector, transforming the continuous time frame into a discrete single-frame analysis problem, which is convenient for digital processing.

[0022] Step S2 includes the following sub-steps: Step S21: Perform a Fast Fourier Transform on the time-domain observation vector of each frame to obtain the frequency-domain observation vector. Step S22: Extract the amplitude for each frequency point in the frequency domain observation vector to obtain the frequency domain amplitude sequence, and extract the phase of each frequency point as the original phase and save it to the phase buffer.

[0023] In one embodiment of the present invention, the time-domain observation matrix obtained in step S1 has L=400, corresponding to a 25ms frame length and a sampling rate of 16kHz, and K=100, representing a speech segment of approximately 2.5 seconds. The t-th frame vector in the time-domain observation matrix is ​​the time-domain observation vector. The frequency domain observation vector is obtained through Fast Fourier Transform processing. ,in, , After completing the Fast Fourier Transform, each frequency point of the frequency domain observation vector is deconstructed to separate the amplitude sequence and original phase required for subsequent processing. The square root of the sum of the squares of the complex number and the imaginary part of each complex number in the frequency domain vector is calculated to obtain the amplitude at that frequency point. Finally, the amplitudes of all 512 frequency points are arranged in order to form a frequency domain amplitude sequence, used to intuitively describe the energy distribution remaining in each frequency band of the current frame. For example, a vowel frame is plotted to show a series of peaks at its fundamental and harmonic frequencies. The original phase is calculated by the phase angle corresponding to each complex number and stored in a time-domain phase buffer. The phase buffer is set as a circular queue structure, storing the original phases of the most recent frames (e.g., 20 frames, 400 milliseconds). When new phase data is stored, it overwrites the oldest frame data, preserving the density of historical phases. When the signal needs to be modified or reconstructed later, the phase continuity between adjacent frames can be used to ensure that the synthesized sound is smooth and natural, avoiding metallic sounds or jitter distortion. This embodiment utilizes the Fast Fourier Transform algorithm to convert the waveform signal representing the change of sound over time into a spectral signal that reveals its inherent frequency components, effectively improving the flexibility of noise reduction and enhancement.

[0024] Step S3 includes the following sub-steps: Step S31: The minimum value tracking method is used to calculate the observation noise of the frequency domain amplitude sequence to obtain the observation noise variance; Step S32: Calculate the objective acoustic metrics for the current frame. The objective acoustic metrics include full-band signal-to-noise ratio, frequency band signal-to-noise ratio, noise stationarity measure, noise burst ratio, and estimation error statistics. Step S33: The objective acoustic indicators are dynamically calculated through a hierarchical decision-making mechanism to obtain the sensitive parameters of the current frame; Step S34: Perform first-order low-pass filtering smoothing on the sensitive parameters of the current frame and output the risk sensitive parameters.

[0025] In this embodiment, the full-band signal-to-noise ratio (SNR) is used to calculate the energy ratio of signal to noise across the entire audible frequency range, providing a macroscopic assessment of the SNR level. The value ranges from -10 dB (extremely noisy) to 30 dB (extremely quiet). The expression for calculating the full-band SNR is: ; in, Signal-to-noise ratio (SNR) across the entire frequency band, expressed in decibels. For frequency point index, Let k be the power at frequency k in frame t. Let V be the observation noise variance at the k-th frequency point; Bandwidth signal-to-noise ratio (SNR) is used to divide the frequency spectrum into 24 sub-bands according to the characteristics of human hearing, and calculate the SNR of each sub-band. Bandwidth SNR reveals the non-uniformity of noise distribution across frequencies. For example, it can distinguish between low-frequency traffic noise and high-frequency wind noise. The expression for calculating bandwidth SNR is: ; in, Let m be the signal-to-noise ratio of the m-th subband. Let m be the set of frequency points of the m-th sub-band, where m is the sub-band index; The noise stationarity measure quantifies the temporal stability of noise by comparing the degree of change in current noise estimates with historical noise estimates. Completely stationary noise approaches 0, while drastically changing noise approaches 1. The expression for calculating the noise stationarity measure is: ; in, Let t be the noise stationarity measure for the t-th frame. Let Variance be the observation noise variance at the k-th frequency point in the (t-10th)-th frame. For a very small positive number, take ; Noise burst ratio is used to detect brief, intense noise events, such as the sound of a door closing or an object falling. Estimation error statistics are used to monitor the health of the minimum tracking algorithm, counting the proportion of frequency points where actual observations are lower than the estimated noise floor. Estimation error statistics reflect the reliability of the underlying estimation algorithm. The expression for calculating the noise burst ratio is: ; in, Let be the noise burst ratio of the t-th frame. Let this be a counting function, to count the number of frequency points that satisfy the conditions. The standard deviation of the noise at the k-th frequency point; Estimation error statistics are used to monitor the health of the minimum tracking algorithm. They count the proportion of frequency points where the actual observed value is lower than the estimated noise floor, reflecting the reliability of the underlying estimation algorithm. The calculation expression for estimation error statistics is as follows: ; in, Statistics on the estimation error of frame t. For indicator functions, The amplitude at the k-th frequency point of the current frame. This is an estimate of the minimum noise value at the k-th frequency. The hierarchical decision-making mechanism adopts a three-tiered progressive structure, including the first tier of basic risk assessment, the second tier of frequency band equalization burst analysis, and the third tier of dynamic synthesis of sensitive parameters. First-level basic risk assessment: The full-band signal-to-noise ratio (SNR) is divided into three preset risk ranges. In this embodiment, these are set as high-risk, medium-risk, and low-risk environments. When the SNR is below 0 dB, it is determined to be a high-risk environment, indicating that the noise intensity has covered or exceeded the speech signal. When the SNR is between 0 and 10 dB, it is determined to be a medium-risk environment, indicating that the speech is indistinguishable but significantly interfered with by noise. When the SNR is above 10 dB, it is determined to be a low-risk environment, indicating that the speech is dominant. A noise stationarity metric is introduced as a dynamic adjustment factor. If the dynamic adjustment factor exceeds an empirical threshold, such as 0.3, it serves as the stationarity metric threshold. This metric is calculated based on the rate of change of the noise variance. Under normal circumstances, for stationary noise like white noise, this value is close to 0. For completely non-stationary noise such as speech bursts or sudden impacts, this value tends to 1, indicating that the noise is in a non-stationary or rapidly changing state, at which point the uncertainty of the environment increases. The decision-making mechanism will raise the basic risk level obtained in the previous step by one level and output the basic risk level, which is divided into three levels: low, medium, and high. The following table compares the empirical thresholds, i.e., the thresholds for measuring stationarity:

[0026] The table above shows that when the empirical threshold is too low, the system is oversensitive and prone to misjudging stationary noise; when the threshold is too high, it is difficult to identify environmental changes in a timely manner, reducing the risk adjustment capability. Therefore, this embodiment uses 0.3 as the empirical threshold. PESQ is a modified speech quality evaluation, used to simulate the subjective hearing of the human ear. When the PESQ value is 1.0~1.5, the speech quality is very poor; when it is 1.5~2.0, the speech quality is poor; when it is 2.0~3.0, the speech quality is average; when it is 3.0~3.5, the speech quality is good; and when it is 3.5~4.5, the speech quality is excellent. Second-layer band equalization burst analysis: Extract the signal-to-noise ratio (SNR) and noise burst ratio of each frequency band, and calculate the range of SNR for all sub-bands. If the range exceeds a set threshold, it is determined that the current noise is severely unevenly distributed in the frequency spectrum, indicating the presence of some prominent narrowband interference or steady-state noise at a specific frequency. In this case, the frequency band imbalance is marked as true. Then, the burstability in the time domain is analyzed, and the noise burst ratio is checked. For example, 15 dB represents the frequency band coverage of human hearing. If the SNR difference between each frequency band exceeds 15 dB, it means that the noise exhibits narrowband characteristics. This noise is extremely uneven in frequency, and 15 dB represents a relatively large dynamic range. This is sufficient to eliminate natural fluctuations caused by everyday speech or slight background noise, ensuring that the frequency band adjustment mechanism is triggered only when a specific interference source appears. If the noise burst ratio exceeds a certain threshold, it is determined that there is a significant burst noise event in the current environment, such as impact sound, keyboard sound, etc. At this time, the burst noise flag is marked as true, and this flag will trigger a special protection mechanism for transient interference: for example, 0.1. If more than 10% of the frequency points in a frame of signal show a sudden surge in energy, statistically it is highly likely that it is not caused by normal speech, but by external physical impact or equipment operation sound. Dynamic synthesis of sensitive parameters in the third layer: Based on the basic risk level output from the first layer, a basic parameter value is directly mapped to it. Low-risk environments are mapped to low values, such as 0.2, for light processing; medium-risk environments are mapped to medium values, such as 0.5, for balancing noise reduction and distortion; and high-risk environments are mapped to higher values, such as 0.8, for deep noise reduction. Next, frequency band equalization adjustment is performed. If the frequency band imbalance flag of the second layer output is true, the use of unified basic parameters is abandoned. Instead, based on the signal-to-noise ratio of each sub-band, the sub-sensitivity parameters of each frequency band are calculated independently through the Sigmoid function. This function is designed to output high parameter values ​​(strong suppression) when the sub-band signal-to-noise ratio is low and low parameter values ​​(weak suppression) when the sub-band signal-to-noise ratio is high, so as to achieve fine, frequency adaptive control. If the burst noise flag in the second layer output is true, a fixed offset (e.g., 0.2) is added to the current base parameter value. An offset of 0.2 is sufficient to significantly change the gain coefficient of the filter, thereby quickly suppressing burst pulses in the next few frames. At the same time, this value will not be so large as to cause the audio signal to appear as if it is completely cut off momentarily. However, it will not exceed the allowed upper limit (e.g., 0.9, 0.9 leaves a 10% soft buffer to ensure that even under extreme risk, the system still retains the most basic speech fundamental frequency characteristics, ensuring that the call is still identifiable despite distortion, and preventing the system from diverging). This is used to temporarily enhance the suppression power of the noise reduction algorithm to quickly suppress burst interference and prevent it from causing auditory discomfort or damaging the speech structure. An estimation error statistic is introduced as a confidence feedback. The estimation error statistic reflects the reliability of the underlying noise estimation algorithm. If the error statistic is high, it indicates that there may be a deviation in the current noise estimation. At this time, the sensitive parameter obtained in the previous step will be reduced proportionally to reduce the risk of overly aggressive decision-making due to unreliable input data and enhance the overall robustness of the system. The aforementioned qualitative judgments and identifications are quantified into a specific risk sensitivity parameter between 0 and 1. The higher the risk sensitivity parameter, the higher the environmental risk; the lower the risk sensitivity parameter, the more environmentally friendly. The key purpose of this embodiment is that the CRSFIR filter provides a risk-sensitive parameter θ, which can reflect the severity of the acoustic environment and noise characteristics in real time. Traditional noise reduction algorithms typically use fixed parameters, making it difficult to achieve optimal results under both stationary and sudden noise conditions. This step, through quantitative evaluation of multiple dimensions such as noise stationarity, spectral imbalance, and estimation algorithm reliability, and through a three-layer progressive decision-making process, enables θ to distinguish between subtle environmental differences such as low signal-to-noise ratio but stationary noise and non-stationary noise containing sudden interference. This guides the filter to adjust its robustness in step S4: when θ increases, the filter becomes more conservative, prioritizing the suppression of abnormal fluctuations (sudden noise); when θ decreases, the filter is closer to the minimum mean square error estimate, preserving more speech details. Furthermore, the introduction of estimation error statistics forms a self-monitoring mechanism, preventing risk parameter errors caused by noise estimation failures, further improving the overall robustness of the system.

[0027] Step S4 includes the following sub-steps: Step S41: Extract the frequency domain amplitude sequence of the most recent N frames of the current frequency point from the amplitude buffer and use it as the observation vector; Step S42: Define a state scalar for the observation vector corresponding to each frequency point, and establish a state space model to obtain state space data; Step S43: Input the state space data into the CRSFIR filter to calculate and obtain the instantaneous state estimation sequence; Step S44: Construct an extended matrix based on the state-space parameters, risk-sensitive parameters, and observation sequence; construct a system of linear equations based on the extended matrix and instantaneous state estimates; and calculate the amplitude estimate at the current frequency point. Step S45: Repeat steps S42, S43 and S44 for all frequency points, arrange the obtained amplitude estimates of each frequency point according to frequency, and combine them into the first estimated amplitude spectrum.

[0028] First, by extracting multiple frames of historical observation data from the phase buffer, a temporal correlation between frequency points is established. Traditional methods only analyze the information of the current frame, while this method utilizes continuous observations of the most recent N frames to form an extended observation vector in the time dimension, which can distinguish between steady-state noise and transient speech components: noise exhibits strong temporal correlation, while speech presents specific time-varying patterns. For example, for continuous background fan noise, multi-frame observations show that its amplitude changes gradually; for speech plosives, it exhibits sudden spikes. This significantly improves the separability of noise and speech. Each frequency point is defined as a state scalar with an internal state. Its evolution over time is described by a state equation, and the relationship between the state and the actual measured value is established by an observation equation. As one of the innovations of this invention, compared with the traditional Kalman filter, the CRSFIR filter avoids the state drift problem by means of constraints while maintaining the efficiency of recursive calculation. It generates the current estimate by weighting and combining multiple historical states. The weight coefficients are dynamically adjusted according to the signal characteristics. The constraints of the filter ensure that the estimated value does not exceed a reasonable physical range. This is especially important when processing weak speech signals in strong noise environment, and effectively prevents the estimation from collapsing. The construction of the extended matrix and the linear equation system enables the fusion and optimization of multi-source information. It integrates state-space parameters, risk-sensitive parameters, and actual observation sequences into a unified mathematical framework. The optimal amplitude estimate is obtained by solving the linear equation system. The state parameters provide prior information, the observation sequence provides likelihood information, and the risk parameter adjusts the conservatism of the estimate. When the risk-sensitive parameter is high, it relies more on the prediction of the state model; when the risk parameter is low, it trusts the observation data more. This adaptive trade-off mechanism enables the system to automatically adjust the estimation strategy under different signal-to-noise ratio conditions.

[0029] By processing all frequency points in parallel, a complete first estimated amplitude spectrum is generated. This amplitude spectrum has effectively suppressed non-stationary noise and sudden interference, while preserving the speech harmonic structure. It can be directly used for speech activity detection and usage classification, or combined with the saved original phase to reconstruct the enhanced time-domain speech signal through inverse FFT.

[0030] Compared with traditional methods, the first estimated amplitude spectrum has higher time-frequency resolution, enabling more precise differentiation of speech harmonic structure and noise components; it has smaller estimation bias, maintaining the integrity of key speech features even under low signal-to-noise ratio conditions; and it has weaker edge effects, resulting in smoother and more natural transitions between adjacent frames.

[0031] Step S42 also includes the following sub-steps: Iterate through each frequency point and, based on the observation vector obtained in step S41, establish a state-space model using a first-order Markov transform. The calculation expression for the state-space model is as follows: ; ; in For the first Time and Frequency The state scalar These are the state transition coefficients. Setting the state transition coefficients to constants close to 1 (e.g., 0.98) indicates that the speech amplitudes of adjacent frames are highly correlated. The process noise is Gaussian distributed. For the first Time and Frequency The observation scalar, H is the observation coefficient (set to 1). The observation noise is Gaussian distributed. Based on the state equation of the state scalar, the initial state estimate is calculated using the least squares estimation method. The initial state estimate and the observation sequence are then combined. The parameters corresponding to the state space model are input into step S43 as state space data, and the observation sequence is used to represent the N latest state estimation observations from the m-th frame to the k-th frame.

[0032] In this embodiment, the closer the state transition coefficient F is to 1, the smoother the filter's response to state changes, but the ability to track rapidly changing speech is slightly weaker; conversely, the tracking ability is enhanced, but the estimation variance may increase. The initial value or proportional relationship between the process noise variance and the observation noise variance is derived and set based on the observation noise variance estimated in step S31. The initial value of the observation noise variance is set to be proportional to the observation noise variance, and the process noise variance is set to a small fraction (such as 0.001 to 0.01 times) of the initial value of the observation noise variance to reflect the situation where the process noise is much smaller than the observation noise. The calculation expression for the least squares estimation method is as follows: ; in This is the initial state estimate. To expand the observation matrix, based on the latest N frames of observation data, the state transition coefficients (F) and the statistical characteristics of noise in each frame are considered. By deducing the true speech amplitude of the m-th frame, The best linear unbiased estimate; By employing least squares estimation on N frames of data containing future information, the state transition coefficients and noise statistics are utilized. Optimal weighted fusion was performed, and the weighted least squares estimation is the best linear unbiased estimate under the Gaussian assumption. Given a model and statistically accurate noise, the given initial state estimate has the smallest variance among all linear estimates, providing a high-precision starting point for the entire filtering process.

[0033] Step S43 also includes the following sub-steps: Set the accumulation interval to a reasonable range within the observation window: Preset accumulation interval length The total window length is 70% of the total window length, which is the length of N frames. The start time s and end time k-1 of the accumulation interval are set. When the start time s is greater than or equal to m, the accumulation interval is within the observation window and the current accumulation interval is a valid interval. When the start time s is less than m, the accumulation interval is not within the observation window and the accumulation interval is converted into an interval with m as the start time and k-1 as the end time. For each time interval l, according to the cumulative interval length, extract the sub-observation vector and the corresponding observation noise variance vector from the observation sequence, and calculate the gain vector using the instantaneous risk-sensitive filter. The expression for calculating the gain vector is as follows: ; in, This represents the gain vector of the instantaneous risk-sensitive filter for the current accumulation interval length. These are the state transition coefficients. As a risk-sensitive parameter, This represents the average value of the observed noise variance vector; The instantaneous state estimate is calculated based on the gain vector. The expression for calculating the instantaneous state estimate is as follows: ; in, From frame m to the first The sum of the observed amplitudes of the frames, i.e. , This represents the gain vector of the instantaneous risk-sensitive filter for the current accumulation interval length. This is an instantaneous state estimate; Will Write the first estimated amplitude spectrum at the corresponding position, and after traversing all frequency points, you will obtain the complete first estimated amplitude spectrum of the current frame.

[0034] By presetting the cumulative interval length and setting an effectiveness detection mechanism, data-driven time-scale adaptability is achieved. When the starting time s < m, the interval is automatically adjusted to [m, k - 1], ensuring that the filter always calculates based on the currently available and latest observation data. Therefore, when there is a noise burst or a drastic change in speech activity, the observation window can be quickly shrunk to focus on recent data, avoiding the estimation bias caused by using outdated historical data in traditional methods. The gain vector nonlinearly couples the risk-sensitive parameter and the observation noise variance. When the risk-sensitive parameter is larger, the item increases significantly, resulting in a decrease in the gain vector value, causing the filter to reduce the trust weight for the sub-observation vector and instead rely more on the model prior (state transition prediction). Therefore, in a strong noise or non-stationary environment, it can effectively suppress the contamination of abnormal observation values to the state estimation, prevent the filter from diverging, and improve the robustness of the system;

[0035] Therefore, as shown in the table, as the cumulative interval increases, the filtering stability increases, but the delay synchronization rises; after 70%, the performance improvement slows down. Therefore, this embodiment adopts a cumulative interval length of 70%. ]

[0036] Step S44 specifically includes: Extract the instantaneous state estimation sequence. The risk-sensitive parameter, sub-observation vector, and parameters of the state space model are used to construct the extended matrix corresponding to the current frequency point f. The extended matrix is a symmetric real matrix with a dimension of . The construction method of the extended matrix is: The extended matrix is divided into sub-matrix blocks, including the upper-left sub-matrix block, the middle sub-matrix block, and the lower-right sub-matrix block. The upper-left sub-matrix block represents the constraint relationship between the system state equation, process noise, observation noise, and risk sensitivity. The middle sub-matrix block and the lower-right sub-matrix block reflect the coupling relationship between the instantaneous state estimation sequence, the current observation vector, and the risk-sensitive parameter; The instantaneous state estimation sequence and the observation vector at the current moment are concatenated to obtain the right-side vector. The row indices of the extended matrix are divided into the first index set and the second index set. The first index set is the set of the global state and the current amplitude estimation , and the second index set is the set of historical intermediate variable indices . According to the first index set and the second index set, the right-side vector is divided into the first right-side sub-vector and the second right-side sub-vector. The first right-side sub-vector is the relevant component corresponding to the first index set, and the second right-side sub-vector is the relevant component corresponding to the second index set; Extract submatrices based on the first and second index sets. The submatrices include a first submatrix, a second submatrix, a third submatrix, and a fourth submatrix. The first submatrix is ​​used to extract elements whose row index and column index are both in the first index set. The second submatrix is ​​used to extract elements whose row index is in the first index set and whose column index is in the second index set. The third submatrix is ​​the transpose of the second index set. The fourth submatrix is ​​used to extract elements whose row index and column index are both in the second index set. Based on the corresponding blocks of the right-hand vector, the reduced coefficient matrix and the low-dimensional linear equation system are constructed using the Schur complement form. The calculation expressions for the reduced coefficient matrix and the low-dimensional linear equation system are as follows: ; ; ; in, This is the first submatrix. For the second submatrix, It is the inverse of the fourth submatrix. For the third submatrix, The first right sub-vector, The second right sub-vector, Let A be the parameter vector to be solved, and let A be a real symmetric matrix. The result vector; The amplitude estimate of the current frequency point f at time k is obtained by solving the low-dimensional linear equation system.

[0037] In this embodiment, the extended matrix has a dimension of The structure of the top-left submatrix block of the real symmetric matrix is ​​as follows: Noise penalty used to encode state transition constraints and risk perception. Used for regularization of initial state uncertainty, where Q is the reciprocal of the process noise variance q. Negative correlation constraints representing state transitions. Used to represent the joint constraint of state evolution and observation noise, where R is the reciprocal of the variance of observation noise; Among them, the first of the extended matrix Line number The column elements represent the joint penalty term of the observation constraint and risk sensitivity at the current moment; Line 1, page The elements of the column are This reflects the decoupling relationship between the initial state and the current observation, in the second row (...). The elements of column ) are This reflects the correlation constraints between state prediction and current observation; The row index of the intermediate submatrix block is from 3 to Column index from arrive ,when At that time, it represents the constraint between the first instantaneous estimate and the system state; By constructing a maximum a posteriori probability estimation problem with risk-sensitive parameters and applying the Schur complement method for intelligent dimensionality reduction, a high-dimensional optimization problem is cleverly transformed into a solution of an extremely low-dimensional linear equation system. This ensures the optimality of the estimation and avoids the numerical instability risk brought about by high-dimensional matrix operations. As a result, the final output frequency amplitude estimate, under the given model and data conditions, reaches the theoretical optimum or near-optimum in the sense of minimum mean square error, providing a solid mathematical guarantee for the ultimate improvement of voice quality. Through the Schur complement dimensionality reduction strategy, the complexity of the solution is drastically reduced from being related to the length of history to the complexity of only needing to process a fixed 3×3 equation system. Although it is necessary to calculate the inverse of an intermediate matrix in the early stage, this matrix is ​​a diagonal matrix, and the inversion is efficient. This embodiment enables the theoretically complex optimal estimation to run stably and in real time on embedded devices such as Bluetooth headsets where computing resources, storage space, and power consumption are extremely limited, with extremely low frame processing latency.

[0038] Step S5 includes the following sub-steps: Step S51: Perform speech activity detection on the first estimated amplitude spectrum to obtain speech detection results. The speech detection results include speech frames and non-speech frames. Speech activity detection includes calculating the frequency band energy of the first estimated amplitude spectrum for each frame, dividing the frequency band energy according to a preset speech threshold, marking the frame corresponding to the frequency band energy as a speech frame when the frequency band energy of the existing frame is greater than the preset speech threshold, and marking the frame corresponding to the frequency band energy as a non-speech frame when the frequency band energy of the existing frame is less than the preset speech threshold, and smoothing the speech detection results to obtain a speech tag sequence. Step S52: Set a time window and extract user temporal features based on speech tag sequences. User temporal features include speech activity, average speech segment length, average silence segment length and number of speech segments, mean energy change rate and variance of energy change rate. Speech activity is the ratio of the number of speech frames in the current time window to the total number of speech frames, used to characterize the overall density of speech use in this time period. The processing logic for the average speech segment length is to extract a continuous sequence of speech markers from speech frames as speech segments, and calculate the average length of each speech segment to obtain the average speech segment length, used to characterize the typical level of a user's single speaking duration. The processing logic for the average silence segment length is to extract a continuous sequence of speech markers from non-speech frames as non-speech segments, calculate the average length of all non-speech segments to obtain the average silence segment length, used to characterize the typical interval of pauses between speech segments, and count the number of speech segments in the current time window to obtain the number of speech segments. For frames marked as speech frames, the energy change rate between adjacent speech frames is calculated through the first estimated amplitude spectrum to obtain the mean and variance of the energy change rate, which are used to reflect the fluctuation and stability of the speech energy when the user speaks. Step S53: Based on user temporal characteristics and the mean and variance of energy change rates, pattern recognition is performed to obtain the final usage results. The final usage results include continuous dialogue communication, silent listening, and discrete interaction. Set a first threshold, a second threshold, and a third threshold. The first threshold is used to determine speech activity, the second threshold is used to determine the average silent segment length, and the third threshold is used to determine the average speech segment length. When the voice activity level is greater than or equal to the first threshold, the average silence segment length is less than the second threshold, and the average voice segment length is greater than or equal to the third threshold, the current time window is determined to be a continuous dialogue communication situation. When the voice activity level is less than the first threshold and the average length of the silent segment is greater than or equal to the second threshold, the current time window is determined to be a silent listening situation. The remaining time window that does not meet the conditions of continuous dialogue communication and silent listening is determined to be a discrete interaction situation.

[0039] In this embodiment, the rate of energy change is defined as the absolute value of the logarithmic difference between the total energy of two adjacent speech frames. : ; in, The energy of the first estimated amplitude spectrum at time t; statistical analysis of all values ​​within this time window. The mean and variance; The preset speech threshold is 0.15, the smoothing window length is 3 frames, the shortest speech segment is 2 frames, the shortest silence segment is 2 frames, the preset time window length is 5 seconds, the first threshold is 0.6, representing speech activity, calculated as the ratio of the number of speech frames to the total number of frames, the second threshold is 0.75, and the third threshold is 0.5. The units for the second and third thresholds are seconds. The settings are determined as follows:

[0040] Therefore, when the first threshold is set to 0.6 seconds, the second threshold is set to 0.75 seconds, and the third threshold is set to 0.5 seconds, the comprehensive classification accuracy for the three scenarios of continuous dialogue communication, silent listening, and discrete interaction is the highest, and thus this is the preferred parameter combination in this embodiment. By intelligently categorizing complex voice interaction modes into three core scenarios—continuous dialogue communication, silent listening, and discrete interaction—it is easier to implement differentiated resource scheduling strategies for different scenarios. For example, when the user is in a continuous dialogue scenario, the stability of the communication link and the noise reduction intensity are automatically enhanced. In a silent listening scenario, audio playback quality is prioritized and battery life is optimized to the extreme. In a discrete interaction scenario, a balance needs to be struck between wake-up and power consumption. This effectively solves the pain points of traditional headphones having limited functionality and fixed settings, and improves the overall user experience satisfaction in different scenarios.

[0041] To verify the parallel message of the present invention, an objective evaluation was conducted under a simulated typical call environment. The test platform was a commercial Bluetooth headset development board (ARM Cortex-M4 core, 200MHz main frequency, 16kB L1 cache), with a signal sampling rate of 16kHz, a frame length of 25ms (400 points), a frame shift of 10ms (160 points), and noise selected from the NOISEX-92 database for babble noise and non-stationary street noise. The signal-to-noise ratio was set to 5dB.

[0042] The table below compares voice enhancement performance:

[0043] A higher PESQ score is better (PESQ stands for Perceived Voice Quality Assessment). Noise suppression refers to the decibel attenuation of output and input noise power. This invention achieved the highest PESQ and the largest noise suppression under both types of non-stationary noise, while the processing time per frame was only 1.2ms, which fully meets the latency requirements of real-time calls.

[0044] The table below shows the adaptive effect of the risk-sensitive parameter θ (street noise, SNR=5dB):

[0045] Speech distortion is measured by cepstral distance. The smaller the value, the lower the distortion. The lower the noise residue, the more thorough the noise reduction. The adaptive θ strategy maintains strong noise suppression capabilities close to those of a high fixed θ while achieving significantly lower speech distortion than a fixed high θ. This demonstrates that the risk-sensitive parameters generated by the hierarchical decision-making mechanism can effectively balance noise reduction robustness and speech fidelity.

[0046] Example 2, refer to Figure 2 It provides a Bluetooth headset usage analysis system based on big data, including a data acquisition module, a frequency domain analysis module, a risk-sensitive parameter estimation module, an amplitude estimation module, and a usage analysis module; The acquisition module is used to acquire multi-channel speech signals containing environmental noise and residual noise through the built-in microphone array of the headset, and to acquire the speaker reference signal. The multi-channel speech signal and the reference signal are converted from analog to digital to obtain a discrete time series. The reference signal is acoustically echo-cancelled using an adaptive filter to obtain a multi-channel de-echo signal. The de-echo signal is beamformed to obtain a single-channel enhanced signal, and the time-domain observation matrix is ​​constructed by windowing the frame. The frequency domain analysis module performs a fast Fourier transform on the time domain observation vector of each frame to obtain the frequency domain observation vector. It extracts the amplitude of each frequency point to form a frequency domain amplitude sequence and extracts the corresponding phase information. The original phase of each frequency point is saved to the phase buffer. The risk-sensitive parameter estimation module is used to estimate the observation noise variance of the frequency domain amplitude sequence using the minimum value tracking method, calculate the objective acoustic indicators of the current frame, including full-band signal-to-noise ratio, frequency band signal-to-noise ratio, noise stationarity measure, noise burst ratio, and estimation error statistics. Through a hierarchical decision mechanism, the objective acoustic indicators are mapped to the sensitive parameters of the current frame, and the sensitive parameters are smoothed by first-order low-pass filtering to output the risk-sensitive parameters. The amplitude estimation module is used to extract risk-sensitive parameters, frequency domain amplitude sequences, and original phases, and calculate the first estimated amplitude spectrum through frequency domain amplitude state estimation processing of CRSFIR. The analysis module is used to detect and classify speech activity based on the first estimated amplitude spectrum to obtain the final usage results.

[0047] In this embodiment, based on the adaptive estimation of risk-sensitive parameters and the effective application of CRSFIR filters, the risk-sensitive parameter estimation module breaks free from the limitations of traditional fixed parameter settings. Through real-time analysis of multi-dimensional acoustic features such as full-band / band signal-to-noise ratio, noise stability, and burst ratio, it dynamically generates risk-sensitive parameters that adapt to the current environment. This enables the system to pursue high-precision estimation under stable noise and automatically enhance robustness under burst noise, thereby improving the speech enhancement quality in complex acoustic environments. By combining the limited memory and accumulation risk sensitivity of the CRSFIR filter, the system can effectively suppress non-stationary noise while avoiding the error accumulation of traditional infinite impulse response filters. This enables the lightweight deployment of speech signal processing technology in Bluetooth headsets and provides the first estimated amplitude spectrum as input for subsequent usage analysis. The analytics module, based on enhanced voice signals, enables deep insights into user behavior patterns, improving user experience and device efficiency.

[0048] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for analyzing Bluetooth headset usage based on big data, characterized in that, Includes the following steps: Step S1: Acquire the original multi-channel speech signal and reference signal and perform preprocessing to obtain the time-domain observation vector; Step S2: Perform a fast Fourier transform on the time-domain observation vector to obtain the corresponding frequency-domain observation vector and extract the amplitude of each frequency point to form a frequency-domain amplitude sequence and the original phase; Step S3: The minimum tracking method is used to estimate the observation noise variance of the frequency domain amplitude sequence, the objective acoustic index of the current frame is calculated, and the objective acoustic index is processed through a hierarchical decision-making mechanism to obtain risk-sensitive parameters. Step S4: Extract risk-sensitive parameters, frequency domain amplitude sequence and original phase, and calculate the first estimated amplitude spectrum through frequency domain amplitude state estimation processing of CRSFIR; Step S5: Based on the first estimated amplitude spectrum, perform speech activity detection and classification to obtain the final usage result.

2. The method for analyzing Bluetooth headset usage based on big data as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: The original multi-channel voice signal is synchronously acquired through the built-in microphone array of the Bluetooth headset, and a reference signal is acquired through the speaker. The original multi-channel voice signal includes environmental noise and residual noise. Step S12: The multi-channel speech signal and the reference signal are digitized by an analog-to-digital converter to obtain a discrete time series. Step S13: Use an adaptive filter to perform acoustic echo cancellation on the digitized reference signal to obtain a multi-channel de-echo signal; Step S14: Based on the discrete time series, spatial filtering is performed on the multi-channel de-echo signal using a beamforming algorithm to obtain a single-channel enhanced signal; Step S15: Overlap the single-channel enhancement signal into frames according to a fixed frame length and frame shift to obtain the enhancement signal of each frame. Multiply the enhancement signal point by point using a window function vector of the same frame length to obtain a time-domain observation vector. Arrange each time-domain observation vector in the order of each frame to obtain a time-domain observation matrix (L*K), where L is the row of the time-domain observation matrix and K is the column of the time-domain observation matrix. The ambient noise is the noise of the current environment that is about to be transmitted into the ear, collected by the reference microphone; The residual noise refers to the residual noise near the inner ear canal of the earphone after noise reduction processing.

3. The method for analyzing Bluetooth headset usage based on big data as described in claim 1, characterized in that, Step S2 includes the following sub-steps: Step S21: Perform a Fast Fourier Transform on the time-domain observation vector of each frame to obtain the frequency-domain observation vector. Step S22: Extract the amplitude for each frequency point in the frequency domain observation vector to obtain the frequency domain amplitude sequence, and extract the phase of each frequency point as the original phase and save it to the phase buffer.

4. The method for analyzing Bluetooth headset usage based on big data as described in claim 1, characterized in that, Step S3 includes the following sub-steps: Step S31: The minimum value tracking method is used to calculate the observation noise of the frequency domain amplitude sequence to obtain the observation noise variance; Step S32: Calculate the objective acoustic metrics for the current frame. The objective acoustic metrics include full-band signal-to-noise ratio, frequency band signal-to-noise ratio, noise stationarity measure, noise burst ratio, and estimation error statistics. Step S33: The objective acoustic indicators are dynamically calculated through a hierarchical decision-making mechanism to obtain the sensitive parameters of the current frame; Step S34: Perform first-order low-pass filtering smoothing on the sensitive parameters of the current frame and output the risk sensitive parameters.

5. The method for analyzing Bluetooth headset usage based on big data as described in claim 1, characterized in that, Step S4 includes the following sub-steps: Step S41: Extract the frequency domain amplitude subsequence of the most recent N frames from the frequency domain amplitude sequence, and use it as the observation vector; Step S42: Define a state scalar for the observation vector corresponding to each frequency point, and establish a state space model to obtain state space data; Step S43: Input the state space data into the CRSFIR filter to calculate and obtain the instantaneous state estimation sequence; Step S44: Construct an extended matrix based on the state-space parameters, risk-sensitive parameters, and observation sequence; construct a system of linear equations based on the extended matrix and instantaneous state estimates; and calculate the amplitude estimate at the current frequency point. Step S45: Repeat steps S42, S43 and S44 for all frequency points, arrange the obtained amplitude estimates of each frequency point according to frequency, and combine them into the first estimated amplitude spectrum.

6. The method for analyzing Bluetooth headset usage based on big data as described in claim 1, characterized in that, Step S42 specifically includes: Iterate through each frequency point and, based on the observation vector obtained in step S41, establish a state-space model using a first-order Markov transform. Calculate the initial state estimate using the least squares estimation method based on the state equation of the state scalar. Then, combine the initial state estimate with the observation sequence. The parameters corresponding to the state space model are input into step S43 as state space data. The observation sequence is used to represent the N latest state estimation observations from the m-th frame to the k-th frame.

7. The method for analyzing Bluetooth headset usage based on big data as described in claim 6, characterized in that, Step S43 specifically includes: Set the accumulation interval to a reasonable range within the observation window: Preset accumulation interval length The total window length is 70% of the total window length, which is the length of N frames. The start time s and end time k-1 of the accumulation interval are set. When the start time s is greater than or equal to m, the accumulation interval is within the observation window and the current accumulation interval is a valid interval. When the start time s is less than m, the accumulation interval is not within the observation window and the accumulation interval is converted into an interval with m as the start time and k-1 as the end time. For each time l, the sub-observation vector and the corresponding observation noise variance vector are extracted from the observation sequence according to the cumulative interval length, and the gain vector is calculated using the instantaneous risk-sensitive filter. The instantaneous state estimate is calculated based on the gain vector, and the calculated instantaneous state estimate is stored in array form according to the time series to obtain the instantaneous state estimate sequence.

8. The method for analyzing Bluetooth headset usage based on big data as described in claim 7, characterized in that, Step S44 specifically includes: Extract the instantaneous state estimation sequence, and construct the extended matrix corresponding to the current frequency point f using risk-sensitive parameters, sub-observation vectors, and parameters of the state-space model. The extended matrix has dimension [missing information]. The symmetric real matrix, the method for constructing the extended matrix is ​​as follows: The extended matrix is ​​divided into sub-matrix blocks, which include a top-left sub-matrix block, a middle sub-matrix block, and a bottom-right sub-matrix block. The top-left sub-matrix block represents the system state equation and the constraint relationship between process noise, observation noise, and risk sensitivity. The middle and bottom-right sub-matrix blocks reflect the coupling relationship between the instantaneous state estimation sequence, the current observation vector, and the risk sensitivity parameter. The right-hand vector is obtained by concatenating the instantaneous state estimation sequence and the observation vector at the current moment. The row indices of the extended matrix are divided into a first index set and a second index set. The first index set is the set of the global state and the current magnitude estimate. The second index set is the historical intermediate variable index set. The right-side vector is divided into a first right-side sub-vector and a second right-side sub-vector according to the first index set and the second index set. The first right-side sub-vector is the relevant component corresponding to the first index set, and the second right-side sub-vector is the relevant component corresponding to the second index set. Submatrices are extracted based on a first index set and a second index set. The submatrices include a first submatrice, a second submatrice, a third submatrice, and a fourth submatrice. The first submatrice is used to extract elements whose row index and column index are both in the first index set. The second submatrice is used to extract elements whose row index is in the first index set and whose column index is in the second index set. The third submatrice is the transpose of the second index. The fourth submatrice is used to extract elements whose row index and column index are both in the second index set. Based on the corresponding blocks of the right-hand vector, the reduced coefficient matrix and the low-dimensional linear equation system are constructed using the Schur complement form. The computational expressions for the reduced coefficient matrix and the low-dimensional linear equation system are as follows: ； ； ； in, This is the first submatrix. For the second submatrix, It is the inverse of the fourth submatrix. For the third submatrix, The first right sub-vector, The second right sub-vector, Let A be the parameter vector to be solved, and let A be a real symmetric matrix. The result vector; The amplitude estimate of the current frequency point f at time k is obtained by solving the low-dimensional linear equation system.

9. The method for analyzing Bluetooth headset usage based on big data as described in claim 8, characterized in that, Step S5 includes the following sub-steps: Step S51: Perform speech activity detection on the first estimated amplitude spectrum to obtain speech detection results. The speech detection results include speech frames and non-speech frames. The speech activity detection includes calculating the frequency band energy of the first estimated amplitude spectrum for each frame, dividing the frequency band energy according to a preset speech threshold, marking the frame corresponding to the frequency band energy as a speech frame when the frequency band energy of the existing frame is greater than the preset speech threshold, and marking the frame corresponding to the frequency band energy as a non-speech frame when the frequency band energy of the existing frame is less than the preset speech threshold. Smooth the speech detection results to obtain a speech tag sequence. Step S52: Set a time window, and calculate the user temporal features based on the normalization of the speech tag sequence. The user temporal features include speech activity, average speech segment length, average silence segment length and number of speech segments, mean energy change rate and variance of energy change rate. The speech activity is the ratio of the number of speech frames in the current time window to the total number of speech frames. The processing logic for the average speech segment length is to extract a continuous speech marker sequence of speech frames as a speech segment, and calculate the average speech segment length by averaging the length of each speech segment. The processing logic for the average silence segment length is to extract a continuous speech marker sequence of non-speech frames as a non-speech segment, calculate the average length of all non-speech segments to obtain the average silence segment length, and count the number of speech segments in the current time window to obtain the number of speech segments. For a frame marked as a speech frame, the energy change rate between adjacent speech frames is calculated using the first estimated amplitude spectrum to obtain the mean and variance of the energy change rate. Step S53: Based on user temporal characteristics and the mean and variance of energy change rates, pattern recognition is performed to obtain the final usage results. The final usage results include continuous dialogue communication, silent listening, and discrete interaction. A first threshold, a second threshold, and a third threshold are set. The first threshold is used to determine speech activity, the second threshold is used to determine the average silent segment length, and the third threshold is used to determine the average speech segment length. When the voice activity level is greater than or equal to the first threshold, the average silence segment length is less than the second threshold, and the average voice segment length is greater than or equal to the third threshold, the current time window is determined to be a continuous dialogue communication situation. When the voice activity level is less than the first threshold and the average length of the silent segment is greater than or equal to the second threshold, the current time window is determined to be a silent listening situation. The remaining time window that does not meet the conditions of continuous dialogue communication and silent listening is determined to be a discrete interaction situation.

10. A Bluetooth headset usage analysis system based on big data, applied in the Bluetooth headset usage analysis method based on big data as described in any one of claims 1-9, characterized in that, It includes an acquisition module, a frequency domain analysis module, a risk-sensitive parameter estimation module, an amplitude estimation module, and a usage analysis module; The acquisition module is used to acquire multi-channel speech signals containing environmental noise and residual noise through the built-in microphone array of the earphone, and to acquire the speaker reference signal. The multi-channel speech signal and the reference signal are converted from analog to digital to obtain a discrete time series. The reference signal is acoustically echo-cancelled using an adaptive filter to obtain a multi-channel de-echo signal. The de-echo signal is beamformed to obtain a single-channel enhanced signal, and the signal is framed and windowed to construct a time-domain observation matrix. The frequency domain analysis module is used to perform a fast Fourier transform on the time domain observation vector of each frame to obtain the frequency domain observation vector, extract the amplitude based on each frequency point to form a frequency domain amplitude sequence, and extract the corresponding phase information; and save the original phase of each frequency point to the phase buffer. The risk-sensitive parameter estimation module is used to estimate the observation noise variance of the frequency domain amplitude sequence using the minimum value tracking method, calculate the objective acoustic indicators of the current frame, including full-band signal-to-noise ratio, frequency band signal-to-noise ratio, noise stationarity measure, noise burst ratio, and estimation error statistics. The objective acoustic indicators are mapped to the sensitive parameters of the current frame through a hierarchical decision mechanism, and the sensitive parameters are smoothed by first-order low-pass filtering to output the risk-sensitive parameters. The amplitude estimation module is used to extract risk-sensitive parameters, frequency domain amplitude sequences, and original phases, and calculate the first estimated amplitude spectrum through frequency domain amplitude state estimation processing of CRSFIR. The usage analysis module is used to detect and classify speech activity based on the first estimated amplitude spectrum to obtain the final usage results.

Images