Two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering, device and storage medium thereof

Abstract

The present invention discloses a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering, as well as its device and storage medium. The method comprises the following steps: performing preliminary noise reduction on the speech signal in the frequency domain by using enhanced spectral subtraction; identifying the state space model parameters of the speech signal based on a linear predictive coding algorithm, and performing secondary noise reduction on the sound signal in the time domain by using a Kalman filter algorithm; and outputting noise-reduced speech after repeating the parameter identification and secondary noise reduction process on the frame-by-frame signal multiple times. The method of the present invention can effectively filter out environmental noise in a complex noise environment and improve the quality of the speech signal. The method is suitable for single-microphone noise reduction of a cochlear implant and can achieve efficient noise reduction in a cochlear implant with limited computing resources, thereby improving the wearer's sound perception ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech signal processing, and in particular relates to a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering, a device thereof, and a storage medium. Background Art

[0002] Cochlear implants are currently the key electronic implants for addressing severe and profound sensorineural hearing loss. Their core function is to convert external speech signals into electrical signals and activate the auditory nerve through electrical stimulation to achieve sound perception. However, in actual application environments, speech signals are often accompanied by complex and variable background noise (such as traffic noise, crowd noise, wind noise, etc.), which will significantly reduce the quality of the speech signal and thus affect the speech intelligibility and auditory comfort of cochlear implant users. Moreover, because cochlear implants are implanted devices, they are required to be as simple and lightweight as possible to be easy for users to carry. Therefore, the development of an efficient, low-complexity speech noise reduction method suitable for cochlear implant systems has important theoretical value and practical application significance.

[0003] Existing speech noise reduction methods primarily include spectral subtraction, Wiener filtering, minimum mean square error estimation, statistical model-based noise reduction algorithms, and Kalman filtering. Spectral subtraction, due to its low computational complexity and simple implementation, is widely used in low-power embedded speech processing devices. However, in practical applications, traditional spectral subtraction is prone to errors when estimating the noise power spectrum, resulting in noticeable "musical noise" artifacts remaining in the denoised speech. This artifact is particularly pronounced in noise environments with strong non-stationary characteristics, seriously affecting speech quality and the auditory experience.

[0004] In contrast, the Kalman filter, as a time-domain optimal estimation algorithm, can fully exploit the temporal correlations of speech signals and, in theory, possesses stronger noise suppression capabilities. However, its performance is highly dependent on an accurate speech state-space model. In particular, due to the influence of noise interference and the non-stationarity of speech signals, the state model parameters often have significant uncertainty, further limiting the effectiveness and stability of the Kalman filter in actual speech noise reduction.

[0005] In summary, how to combine the advantages of frequency domain and time domain methods under the constraints of limited computing resources to construct a noise reduction algorithm that is highly adaptable, stable, and capable of dynamically estimating the speech state model has become a key research direction for improving the speech processing performance of cochlear implants. Summary of the Invention

[0006] In order to deal with the complex background noise that often accompanies sound signals in actual application scenarios and improve the speech intelligibility and auditory comfort of cochlear implant wearers, the present invention proposes a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering, as well as its device and storage medium. The method can effectively filter out environmental noise in complex noise environments, improve the quality of speech signals, and can be implemented in a device to process noisy speech.

[0007] In order to achieve the above effects, the technical solutions adopted by the present invention are as follows:

[0008] A two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering comprises the following steps:

[0009] S1: Preprocess the received noisy speech signal, divide each frame of the speech signal into frequency bands after preprocessing, calculate the variance of each frame of the speech signal after frequency band division, perform logarithmic transformation on the variance of each frequency band to obtain the variance result of each frequency band, and judge whether the current frame is a speech frame or a noise frame based on the variance result of each frequency band and the set threshold;

[0010] If the current frame is a noise frame, a static smoothing strategy is used to estimate the noise power spectrum of the noise frame;

[0011] If the current frame is a speech frame, a dynamic smoothing strategy is used to estimate the speech power spectrum of the speech frame;

[0012] Calculate the posterior signal-to-noise ratio based on the ratio of the speech power spectrum to the noise power spectrum estimated for each frame, substitute the posterior signal-to-noise ratio for each frame into the subtraction factor calculation formula to obtain an adaptive subtraction factor, and calculate the preliminary enhanced speech signal based on the adaptive subtraction factor and the enhancement spectrum subtraction formula;

[0013] S2: Preprocess the initially enhanced speech signal, estimate the model parameters, and establish a model for Kalman filtering to obtain a denoised speech signal.

[0014] In this method, enhanced spectral subtraction is used to perform preliminary noise reduction on speech signals in the frequency domain. Linear predictive coding is used to identify parameters of the speech signal's state-space model, and a Kalman filter is used to perform secondary noise reduction on the sound signal in the time domain. This parameter identification and secondary noise reduction process is repeated multiple times on each frame, ultimately outputting the de-noised speech. This method is suitable for single-microphone noise reduction in cochlear implants, achieving efficient noise reduction in cochlear implants with limited computing resources, thereby improving the wearer's sound perception.

[0015] In step S1, the pre-processing specifically includes: framing, windowing, pre-emphasis and short-time Fourier transform.

[0016] In step S2, the pre-processing specifically includes: framing, windowing, pre-emphasis and short-time Fourier transform.

[0017] In step S1, the frequency band division adopts Bark frequency band division.

[0018] In step S1, the current frame is judged as a speech frame or a noise frame based on the comparison of the variance results of each frequency band with a set threshold, which specifically includes:

[0019] When the variance result of the frequency band is greater than or equal to the threshold, the current frame is judged to be a speech frame;

[0020] When the variance result of the frequency band is less than the threshold, the current frame is determined to be a noise frame.

[0021] In step S1, the threshold is set to -10~0dB.

[0022] In step S2, the initially enhanced speech signal is preprocessed, various model parameters are estimated, and a model is established for Kalman filtering to obtain the noise-reduced speech, which specifically includes:

[0023] The initially enhanced speech signal is preprocessed to obtain a framed speech signal, and the speech signal model parameters are estimated for the framed speech signal using a linear predictive coding method;

[0024] The process noise covariance parameters are estimated by taking the residual of the initially enhanced speech signal and the estimated speech signal model parameters, and the observation noise covariance parameters are estimated by the spectral flatness;

[0025] A state-space model is constructed based on the speech signal model parameters, process noise covariance parameters and observation noise covariance parameters. Kalman filtering is performed on the state-space model to remove residual noise and obtain a denoised speech signal.

[0026] A device for implementing the two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering, comprising:

[0027] Enhanced spectrum subtraction processing module, used to pre-process the noisy speech signal, estimate the noise power spectrum and speech power spectrum, and perform spectrum subtraction processing with an adaptive over-subtraction factor to obtain a preliminarily enhanced speech signal;

[0028] The Kalman filter processing module is used to pre-process the initially enhanced speech signal, estimate model parameters, and establish a model for Kalman filter processing to obtain a noise-reduced speech signal.

[0029] A cochlear implant speech processing device includes a microphone, a memory, and a processor. The microphone is used to collect speech signals from an environment, the memory is used to store programs, and the processor is used to execute the two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering. The method of the present invention is designed to be implemented in a cochlear implant device for daily use. It can effectively improve the clarity and intelligibility of speech signals while meeting the lightweight computing power requirements of the cochlear implant device.

[0030] The cochlear implant speech processing device further includes: a human brain electrode patch, which is connected to the processor, and the processor sends the noise-reduced speech signal to the human brain electrode patch.

[0031] A computer-readable storage medium stores a program, which, when executed by a processor, implements the two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering.

[0032] Furthermore, a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering includes the following steps:

[0033] S1: Enhanced spectral subtraction. The speech signal is first preprocessed and divided into subbands using the Bark scale. The subband variance of each frame is calculated. This subband variance is then converted to the logarithmic domain and compared with a set threshold to distinguish speech frames from noise frames. A differentiation strategy is then used to estimate the noise power spectrum based on the frame type. Subsequently, the posterior signal-to-noise ratio of each speech frame is calculated, and the oversubtraction factor in the spectral subtraction formula is dynamically adjusted. The estimated noise power spectrum is then subtracted from the speech signal based on the adaptive oversubtraction factor according to the spectral subtraction formula to achieve denoising. Finally, post-processing is performed to remove any residual transient noise and avoid over-suppression of speech, resulting in a preliminarily de-noised sound signal.

[0034] S2: Time-Domain Modeling and Kalman Filtering. For the enhanced spectral subtraction speech signal, Linear Predictive Coding (LPC) is used to construct a p-order autoregressive (AR) model for each frame and convert it into a state-space model. The required process noise in the model is calculated based on the residuals of the autoregressive model, and the observation noise is calculated using the spectral flatness method. Based on this model, the Kalman filter algorithm is applied to achieve time-domain noise reduction. Within each frame, LPC parameter estimation and Kalman filtering operations are repeated, iteratively updating the model parameters and filtering results, ultimately outputting a stable, pure speech signal.

[0035] Compared with the existing methods, the advantages of the present invention are mainly reflected in:

[0036] 1. The present invention proposes a two-stage speech denoising method that integrates enhanced spectral subtraction and iterative Kalman filtering, which fully combines the processing advantages of frequency domain and time domain to achieve higher quality speech signal denoising.

[0037] Second, in the frequency domain, the enhanced spectral subtraction process dynamically adjusts the over-subtraction factor by introducing a posteriori signal-to-noise ratio estimation, effectively alleviating the "music noise" artifact problem in traditional spectral subtraction and improving the accuracy of noise power spectrum estimation. The spectral subtraction intensity is enhanced in noise frames and weakened in speech frames, reducing distortion in speech segments and improving the human auditory experience.

[0038] 3. Kalman filtering relies on mathematical modeling of the system dynamic process and observation process. By enhancing the coarse noise reduction of the spectral subtraction process, the noise interference in the Kalman filtering process can be reduced, the speech accuracy can be improved, and the error can be reduced.

[0039] 4. In the time domain stage, by combining the LPC-based state space model with the Kalman filter, a signal processing framework capable of adaptive updates is constructed, which significantly improves the modeling and noise reduction capabilities of speech signals under non-stationary noise.

[0040] 5. Iteratively executing modeling and filtering operations effectively reduces the fluctuation of noise reduction performance caused by model uncertainty and improves the stability and accuracy of parameter estimation.

[0041] 6. The method of the present invention has low computational complexity and is suitable for embedded auditory devices such as cochlear implants that have high requirements for power consumption and computing power. While maintaining the lightweight design of the system, it significantly improves speech clarity and user auditory comfort. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0043] Figure 2 This is a schematic diagram of the enhanced spectrum subtraction process of the present invention;

[0044] Figure 3 A schematic diagram of the Kalman filter portion of the iterative state modeling process of the present invention;

[0045] Figure 4 This is a simulation effect diagram of synthesizing a speech segment under a 0dB signal-to-noise ratio condition using the method of the present invention;

[0046] Figure 5 This is a simulation diagram of the speech segment synthesized by the traditional method under the condition of 0dB signal-to-noise ratio;

[0047] Figure 6 The method of the present invention is used to synthesize a speech segment under the condition of 0dB signal-to-noise ratio to simulate the speech spectrogram;

[0048] Figure 7 This is a simulation effect diagram of synthesizing a speech segment using the method of the present invention under a 5dB signal-to-noise ratio condition;

[0049] Figure 8 This is a simulation diagram of the traditional method synthesizing a speech segment under a 5dB signal-to-noise ratio condition;

[0050] Figure 9 The method of the present invention synthesizes a speech segment under the condition of 5dB signal-to-noise ratio and simulates the speech spectrogram;

[0051] Figure 10 This is a simulation diagram of the method of the present invention synthesizing a speech segment under a 10dB signal-to-noise ratio condition;

[0052] Figure 11 This is a simulation diagram of the traditional method synthesizing a speech segment under a 10dB signal-to-noise ratio condition;

[0053] Figure 12 The method of the present invention synthesizes a speech segment under the condition of 10dB signal-to-noise ratio and simulates the speech spectrogram;

[0054] Figure 13 This is a simulation diagram of the method of the present invention synthesizing a speech segment under a 15dB signal-to-noise ratio condition;

[0055] Figure 14 This is a simulation diagram of the traditional method synthesizing a speech segment under a 15dB signal-to-noise ratio condition;

[0056] Figure 15 The method of the present invention synthesizes a speech segment under the condition of 15dB signal-to-noise ratio and simulates the speech spectrogram;

[0057] Figure 16 This is a rendering of the noise reduction method according to a specific embodiment of the present invention under a 5dB signal-to-noise ratio condition in an indoor exhibition hall environment;

[0058] Figure 17 This is a rendering of the noise reduction method according to a specific embodiment of the present invention under a 10dB signal-to-noise ratio condition in an indoor exhibition hall environment;

[0059] Figure 18 This is a rendering of the noise reduction method according to a specific embodiment of the present invention under a 15dB signal-to-noise ratio condition in an indoor exhibition hall environment;

[0060] Figure 19 This is a rendering of the noise reduction method according to a specific embodiment of the present invention under a 5dB signal-to-noise ratio condition in an outdoor station environment;

[0061] Figure 20 This is a rendering of the noise reduction method according to a specific embodiment of the present invention under a 10dB signal-to-noise ratio condition in an outdoor station environment;

[0062] Figure 21 This is a rendering of the noise reduction method according to a specific embodiment of the present invention under a 15dB signal-to-noise ratio condition in an outdoor station environment;

[0063] Figure 22 This is a diagram showing the effects of a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering under a 5dB signal-to-noise ratio according to a specific embodiment of the present invention before and after noise reduction.

[0064] Figure 23 This is a diagram showing the effects of a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering under a 10 dB signal-to-noise ratio according to a specific embodiment of the present invention before and after noise reduction.

[0065] Figure 24 This is a diagram showing the effects before and after noise reduction of a two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering under a 15dB signal-to-noise ratio according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, design ideas and technical solutions of the embodiments of the present invention clearer, the present invention is further described below with reference to the accompanying drawings.

[0067] like Figure 1 、 Figure 2 and Figure 3 As shown, the present invention proposes a two-stage speech noise reduction method based on enhanced spectral subtraction and improved Kalman filtering, the method comprising the following steps:

[0068] Step S1: Enhanced spectrum subtraction processing

[0069] The cochlear implant system uses a micro-electro-mechanical system (MEMS) microphone array to collect raw noisy speech signals. After processing through a preamplifier, the signals enter the digital signal processing unit. Signal preprocessing begins with frame segmentation, Hamming window weighting, and first-order pre-emphasis filtering. The time-domain signal is then mapped to the time-frequency domain using a short-time Fourier transform (STFT). Following this conversion, a Bark scale filter bank performs critical band decomposition, dividing the spectrum into multiple subbands. The variance of each Bark subband is calculated, converted to the logarithmic domain, and compared with a pre-set adaptive threshold to distinguish between speech and noise frames. A noise frame contains only noise, while a speech frame contains both speech and a certain level of noise. Different noise estimation methods are used for different frames to estimate the noise power spectrum. Finally, adaptive over-subtraction spectral subtraction based on signal-to-noise ratio (SNR) estimation is used to perform preliminary speech denoising.

[0070] Step S101: Bark subband division

[0071] According to the theory of human auditory frequency band, each frame of the spectrum after short-time Fourier transform is non-uniformly divided, and the linear spectrum of 20Hz~20kHz is mapped to the perception band of Bark scale. The specific conversion formula is:

[0072]

[0073] is the Bark perception frequency corresponding to each frequency point f, in Bark units, The linear frequency unit is Hz (Hertz), and each frequency point in a frame is divided into the corresponding Bark sub-band through the formula.

[0074] Speech feature frequency band merging:

[0075] A merging strategy combines adjacent Bark subbands into a wideband, making it more suitable for speech feature extraction. The main merged frequency bands include the F1 region (300-1000Hz), which carries the first formant and carries the energy of the vowel fundamental frequency; the F2-F3 region (1000-2500Hz), which contains the second and third formants and determines the timbre of vowels; and the high-frequency consonant region (2500Hz-4000Hz), where the energy of voiceless consonants such as / s / and / ʃ / is concentrated. This method maximizes the preservation of the continuous nature of speech formants, reduces the number of subbands, and reduces the amount of variance calculation. The energy variance of the merged subbands is more stable, improving the accuracy of Voice Activity Detection (VAD) under low signal-to-noise ratio conditions.

[0076] Step S102: Frame type determination

[0077] Based on the Bark subbands obtained in the previous step, the variance of each subband of each frame is calculated. At the same time, to prevent the influence of sudden noise, the variance is smoothed. The smoothing formula is as follows:

[0078]

[0079] For the The variance after frame smoothing, is the subband variance calculated before smoothing, is a smoothing factor that controls the weight of historical information and its value ranges from 0.7 to 0.95.

[0080] The smoothed variance is then converted to the logarithmic domain and compared with the adaptive threshold. The adaptive threshold dynamically adjusts the threshold by the long-term average noise variance to avoid misjudgment of the fixed threshold when the noise changes.

[0081] The formula for converting the band variance to logarithmic domain is as follows:

[0082]

[0083] is the variance result after logarithmic domain transformation, To minimize Logarithm operation overflows.

[0084] The long-term average update adaptive threshold formula is as follows:

[0085]

[0086] For each The adaptive threshold of the frame is initially set to a minimum value, and the weights 0.95 and 0.05 indicate that the long-term average is more dependent on historical values, is insensitive to sudden changes, and is suitable for tracking a stable noise floor.

[0087] The final adaptive threshold formula is as follows:

[0088]

[0089] For the The final adaptive threshold for the frame is expressed in dB (decibels), a logarithmic unit for measuring sound intensity. +3 dB is an empirical value, indicating that the speech segment variance must be at least 3 dB higher than the noise (approximately twice the linear energy). The max (-10dB) value sets the lower threshold to prevent it from being too low (for example, when the long-term average noise level is extremely low). Ensure that the threshold does not fall below -10dB to avoid false positives in extreme cases. The final value is the maximum of the two thresholds. In actual processing, the lower threshold can be set between -10dB and 0dB to adapt to different scenarios.

[0090] Then the speech frame is judged. If the value of the smoothed variance of the current frame converted into the logarithmic domain exceeds the set threshold, the frame is judged to be a speech frame. Otherwise, it is a noise frame and the next frame judgment is entered (for example, The noise variance of the frame is calculated in the logarithmic domain to obtain the variance result -11dB, the adaptive threshold of the current frame is -45 dB, according to The formula shows that the current adaptive threshold is -10db. <-10dB, the current frame is a noise frame, if the variance result of the current frame Greater than the adaptive threshold , then the frame is determined to be a speech frame, and the next frame The variance result of the next frame and for dynamic updates).

[0091] Step S103: Noise power spectrum estimation

[0092] After determining the type of each frame, different noise estimation strategies are adopted for different frames.

[0093] For noisy frames, a static smoothing estimation strategy is used to update the noise power spectrum. The formula is as follows:

[0094]

[0095] in, Indicates the The first The noise power spectrum estimate of the frequency point, Indicates the The first The original signal power spectrum of the frequency point, is the smoothing factor, high The value is suitable for environments with high proportion of steady-state noise and voice, slow noise tracking speed, low The value is suitable for non-stationary noise environments and is used to better track noise changes.

[0096] For speech frames, first calculate the posterior signal-to-noise ratio:

[0097]

[0098] For the Frame No. The posterior signal-to-noise ratio of the frequency point is determined by the current Frame No. Signal power spectrum at frequency points than above Frame No. Noise power spectrum at each frequency point Then the Sigmoid function is used to estimate the probability of speech in the sound signal:

[0099] For the The probability of speech existence at the kth frequency point in the frame, where The larger the value, the stronger the voice signal in the audio, and the closer the probability of existence is to 1. Then, based on the probability calculation, the smoothing factor is dynamically adjusted according to the dynamic smoothing strategy. , the formula is as follows:

[0100]

[0101] Finally, the noise power spectrum of the current frame is updated according to the dynamic smoothing factor:

[0102]

[0103] Step S104: Enhanced spectral subtraction processing

[0104] According to the estimated noise power spectrum, the following enhanced spectrum subtraction method is used for noise reduction:

[0105]

[0106] in Indicates the The first The power spectrum of the sound signal after the frequency point enhanced spectrum subtraction processing, is the overreduction factor ( ), used to control the noise suppression strength, is the power law coefficient (usually 1 or 2), which is used to adjust the nonlinear characteristics of spectral subtraction; It is the spectrum lower limit factor, which prevents the power spectrum from being too small or distorted.

[0107] To improve adaptability, Frame, The over-reduction factor of the frequency point Can be dynamically adjusted according to the signal-to-noise ratio:

[0108]

[0109] in , is an adjustable parameter, , , respectively controlling the steepness of the attenuation curve and the position of the attenuation threshold, A larger value of b provides stronger noise suppression, but may cause speech distortion; a too small value may cause residual noise; a larger value of b will over-reduction factor In the high SNR area, it drops later, retaining more noise suppression, but may over-attenuate the voice. A smaller value will do the opposite. The value should be adjusted according to actual needs. The dynamic adjustment mechanism has a weaker suppression on the speech signal in the speech frame, which can protect the speech quality, and can enhance the suppression strength of the noise signal in the noise frame.

[0110] Calculate the oversubtraction factor Required signal-to-noise ratio The posterior signal-to-noise ratio of the speech frame to calculate the probability of speech presence The formula is different:

[0111]

[0112] Dynamically adjust the subtraction factor based on the estimated speech power spectrum and noise power spectrum at each frequency point of each frame .

[0113] After spectral subtraction, post-processing is performed using the local minimum replacement method. The noise amplitude spectrum estimate is subtracted from the noisy signal amplitude spectrum, and the maximum value in the frequency domain is taken as the noise threshold. If the amplitude spectrum of some frequency points in the frame is lower than the noise threshold, these points are considered to contain mainly noise components and are replaced with the minimum amplitude value of the noise in the adjacent frames, thereby eliminating the "music noise" remaining after spectral subtraction as much as possible.

[0114] Step S2: Time domain modeling and Kalman filtering

[0115] For the speech spectrum sequence after preliminary processing by enhanced spectrum subtraction, the LPC algorithm is used to estimate the AR model parameters frame by frame, and a state space model is constructed, and further noise reduction is performed through Kalman filtering.

[0116] Step S201: State space modeling

[0117] The sound signal satisfies the following p-order autoregressive model:

[0118]

[0119] in , Respectively represent The sound signal and the unknown white noise signal are known at all times. , is the autoregressive parameter to be estimated, which can be estimated using LPC technology. It is worth noting that when When it is a single channel signal, the parameter is a scalar, when For multi-channel signals, the parameters is a matrix. For the convenience of expression, it is assumed below is a single-channel signal. , , then the above autoregressive model can be converted into the following state space equation form:

[0120]

[0121] in , , , Respectively represent The system state at each moment, the measured value, the process noise and the measurement noise, is the state transition matrix, is the observation matrix, , They represent the system matrices respectively, specifically:

[0122]

[0123] The order p of the autoregressive model can be set according to the actual application scenario. In the speech coding (LPC) environment, the recommended order is 10 to 16, in speech enhancement or noise reduction scenarios, the recommended order is 8 to 14, for fundamental frequency estimation (Pitch) the recommended order is 4 to 8, and for music signal analysis the recommended order is 20 to 40.

[0124] Step S202: Noise covariance estimation

[0125] First, the autocorrelation function (ACF) is calculated for each frame, and then the power spectral density (PSD) is obtained through Fourier transform. For the main frequency band of human speech [100Hz, 2000 Hz], the PSD in this interval is extracted and calculated based on the spectral flatness (defined as the ratio of the geometric mean to the arithmetic mean):

[0126]

[0127] It is The power value of the kth frequency point in the frame, N is the number of frequency points on the frame spectrum. The spectrum of the noise frame is relatively flat (high spectral flatness), while the spectrum flatness of the speech frame (especially the voiced sound) is low due to the harmonic structure. This characteristic is used to re-divide the signal frame, and the spectrum flatness of each frame is normalized and compared with the set threshold. The normalized threshold is taken as Frames with spectral flatness greater than a set threshold are considered silent frames. The maximum variance of each silent frame in the interval [100 Hz, 2000 Hz] is taken as the value of the observed noise. The observed noise covariance matrix R is the p-order identity matrix multiplied by the maximum variance in the silent frame.

[0128] The process noise covariance matrix Q is obtained by estimating the forecast error variance when the LPC algorithm estimates the AR model parameters:

[0129]

[0130] The variance is White noise, Q is a p-order matrix:

[0131]

[0132] Step S203: Kalman filter process

[0133] 1. Prediction stage:

[0134] Status prediction:

[0135] Forecast error covariance:

[0136] 2. Update phase:

[0137] Kalman gain:

[0138] Status Update:

[0139] Error covariance update:

[0140] For the current The predicted state value at the moment, for The best estimate of time, is the prior variance estimation matrix at time k, and are the posterior covariance matrices at time k and k-1 respectively, is the Kalman gain at time k, I is the identity matrix with the same dimensions as the previous matrix.

[0141] Step S204: Parameter iterative update

[0142] After completing Kalman filtering for each frame, the filtering results are used to re-estimate the AR model parameters, the observation noise covariance matrix R, and the process noise covariance matrix Q, and then Kalman filtering is performed again. After multiple iterative identification filtering, the speech is finally output. The number of iterations here is set according to the requirements.

[0143] Test simulation software:

[0144] Research on speech denoising algorithms was conducted using the MATLAB R2019b simulation platform. Speech test sets were generated under different noise environments using the SNR synthesis module. Experiments comparing the improved speech denoising algorithm with traditional methods were conducted on the same noisy speech samples. A multidimensional evaluation system was constructed using objective evaluation metrics such as STOI, PESQ, fwSNRseg, and LLR. Statistical methods were used to analyze the experimental data for significant differences, verifying the robustness of the improved algorithm in low signal-to-noise ratio environments and the effectiveness of its improved denoising performance.

[0145] Noise Reduction Metrics: Four metrics are used to measure the algorithm's objective noise reduction performance: frequency-weighted segmental signal-to-noise ratio (fwSNRseg), log-likelihood ratio (LLR), short-time objective intelligibility (STOI), and perceptual evaluation of speech quality (PESQ). fwSNRseg is an improved version of the traditional segmental signal-to-noise ratio (SNRseg), incorporating frequency weighting to better align with human perception. The algorithm divides the input speech signal into frames and calculates the SNR for each frame, applying weights to different frequency components. LLR, based on linear prediction analysis, compares the linear prediction coefficients (LPCs) between the original and processed speech, reflecting the degree of distortion in the speech spectral envelope. The practical range is 0-10, with lower values ​​indicating less distortion. STOI is an objective evaluation metric that measures the improvement in speech intelligibility after noise reduction. Its value ranges from 0 to 1, with higher values ​​indicating higher intelligibility and clarity. PESQ is a standard algorithm recommended by the International Telecommunications Union (ITU) for objectively predicting subjective speech quality. It simulates the human auditory system and comprehensively evaluates speech distortion and noise interference. It is currently the most widely used objective speech quality assessment method. The original PESQ model scores range from -0.5 to 4.5, with higher scores indicating better speech quality.

[0146] To evaluate the noise reduction performance of this method in various noise environments, this case used the artificial noise superposition method to generate noisy speech test sets with signal-to-noise ratios of 5dB, 10dB, and 15dB, respectively, for different speaker genders, noise scenarios, and noise types according to the IEC 60268-4 standard. The case speech was obtained from the NOIZEUS speech database. This method was used for noise reduction. To verify the actual effectiveness of the present invention, it was compared with a method that combines basic spectral subtraction based on the minimum noise tracking algorithm and basic Kalman filtering. The denoising ability of this method was verified by comparing waveform graphs and audio graphs. Figures 4 to 15 This is the effect diagram under different human voice, noise environment and signal-to-noise ratio conditions. Figures 4 to 6 The audio segment is a male voice, the noise scene is a street, and the synthesized speech segment has a speech signal-to-noise ratio of 0dB. Figure 4 The waveform diagram shows the speech output results of the noise-containing frequency band under the method of the present invention; Figure 5The method used in the waveform diagram is a basic spectrum subtraction method based on a minimum noise tracking algorithm, the Kalman filter used is a basic Kalman filter, and the method for measuring the noise matrices R and Q is the same as that used in the present invention. Figure 6 The speech spectrogram before and after noise reduction is compared. Figure 4 and Figure 5 It can be seen that the method of the present invention has a good effect on processing human voice background noise. Its ability to suppress noise and preserve the harmonic structure of human voice is superior to traditional spectral subtraction and Kalman filtering algorithms. It has a fast processing speed and good real-time performance. Comparing the audio graphs also clearly shows that the background noise band is mostly filtered out before and after filtering. Figures 7 to 9 The audio segment is a male voice, the noise scene is a train station, and the signal-to-noise ratio of the synthesized speech segment is 5dB; Figures 10 to 12 The audio segment condition is a female voice, the noise scene is a restaurant, and the signal-to-noise ratio of the synthesized speech segment is 10dB; Figures 13 to 15 The audio segment is a female voice, the noise scene is an airport, and the synthesized speech segment has a signal-to-noise ratio of 15dB. The results show that the method proposed in this paper has excellent speech enhancement effects even under different genders and background noise conditions. This method is validated using four metrics: STOI, PESQ, LLR, and fwSNRseg. By comparing the results of the clean audio with the denoised audio after adding noise, the final STOI score for the speech segment can reach 0.7, and in steady-state noise environments with high signal-to-noise ratios, it can reach above 0.8. The PESQ score is also significantly improved, with lower signal-to-noise ratios in the speech segment showing a greater improvement. The LLR score can reach below 1, indicating a significant improvement. The fwSNRseg metric improves across all audio segments, with similar improvements across different signal-to-noise ratios. The waveform results and actual auditory experience demonstrate that the two-stage speech denoising method proposed in this paper, based on enhanced spectral subtraction and Kalman filtering, has significant advantages in filtering out noise and enhancing human voices. Experimental verification has shown that this method performs well in various noise environments and signal-to-noise ratios.

[0147] At the same time, in order to better verify the effectiveness of this method, three implementation examples are selected to process speech signals in different scenarios and different background noises, and verify them under different signal-to-noise ratio conditions:

[0148] Example 1

[0149] The purpose of Example 1 is to study whether the present method can achieve good noise reduction performance in a typical indoor acoustic environment. In this example, a male voice is selected, and the background noise is mainly intercepted from an indoor exhibition hall. The background noise is a fusion of indoor exhibition hall recordings and white noise. The signal-to-noise ratio of the noisy speech is set to 5 dB, 10 dB, and 15 dB, respectively. Each noisy speech segment is played one by one to test the effectiveness of the method.

[0150] Analyzing the results before and after noise reduction using various metrics reveals that, under the conditions of Example 1, the STOI metric can reach 0.80 or above when compared to the signal-to-noise ratio. Data shows that when the signal-to-noise ratio is high, the degree of STOI improvement decreases as the signal-to-noise ratio increases. The LLR metric can reach below 1 when the signal-to-noise ratio is high, with the improvement increasing as the signal-to-noise ratio decreases. The PESQ metric also improves, but at high signal-to-noise ratios, the PESQ value may decrease. The fwSNRseg metric, both before and after audio noise reduction, improves after filtering and noise reduction using this method, with similar differences across all signal-to-noise ratios.

[0151] like Figures 16 to 18 As shown: Figure 16 、 Figure 17 、 Figure 18 The waveform diagrams are respectively of the noise reduction results of the method for the noise-containing frequency under the conditions of signal-to-noise ratio of 5dB, 10dB and 15dB. From the results, it can be seen that the method has significant denoising efficiency in the noise-dominated area under the condition of Example 1, and can also well maintain the original speech waveform in the area where speech energy is concentrated (such as the voiced segment). After actual filtering, the auditory perception can obviously feel that the background noise intensity has been reduced, and the distortion of the speech part after noise reduction is small, and the speech intelligibility is high. Combining objective indicators and subjective evaluation, the method shows a significant noise reduction advantage in the indoor multi-sound source environment of Example 1.

[0152] Example 2

[0153] The purpose of Example 2 is to study whether this method can have good noise reduction performance in a typical outdoor acoustic environment. In this example, female voice audio is selected, the noise is mainly intercepted from the station, the background noise is fused into the outdoor station vehicle noise and white noise, and the signal-to-noise ratio of the noisy speech is set to 5 dB, 10 dB, and 15 dB respectively. Each noisy speech segment is played one by one to test the effectiveness of the method.

[0154] By analyzing the results before and after noise reduction using various indicators, it can be seen that in the environment of Example 2, when the signal-to-noise ratio is high, the STOI indicator can reach 0.75 or above. In the actual noise reduction process, when the signal-to-noise ratio is high, the degree of improvement in STOI will also decrease as the signal-to-noise ratio increases. The LLR indicator results are the same as in Example 1. When the signal-to-noise ratio is high, the LLR value decreases less. The PESQ indicator will also improve, but may decrease when the signal-to-noise ratio is high. The overall audio fwSNRseg indicator will improve after filtering and noise reduction using this method, and the improvement values ​​at various signal-to-noise ratios are not much different.

[0155] like Figures 19-21 As shown: Figure 19 、 Figure 20 、 Figure 21The waveform diagrams are respectively of the noise reduction results of the method for the noise-containing frequency under the conditions of signal-to-noise ratio of 5dB, 10dB and 15dB. From the results, it can be seen that the method has significant filtering efficiency for various noises in the outdoor station environment under the condition of Example 2. After actual filtering, the background noise can be obviously reduced in the auditory perception, and the distortion of the speech segment is small, which proves that the method also has good noise reduction ability in the outdoor environment of Example 2.

[0156] Example 3

[0157] The purpose of Example 3 is to study the target speech extraction and anti-interference performance of this method in a multi-voice mixed scene, while filtering out excess environmental noise. This example selects male voice audio, with background noise primarily consisting of conversational noise and white noise. The signal-to-noise ratio of the noisy speech is set to 5dB, 10dB, and 15dB, respectively. Each noisy speech segment is played back one by one to test the effectiveness of the method.

[0158] Analyzing the results before and after noise reduction using various metrics reveals that, in Example 3, because the focus is primarily on voice recognition, there is a risk of misidentifying ambient human voices as human voices. Therefore, the STOI metric is slightly lower than in the previous two examples, but can still reach 0.85±0.03 under high signal-to-noise ratio conditions. The remaining metrics show the same results as in Examples 1 and 2. LLR, PESQ, and fwSNRseg generally improve after noise reduction using this method, with the improvement being more pronounced at lower signal-to-noise ratios.

[0159] like Figures 22-24 As shown: Figure 22 、 Figure 23 、 Figure 24 The waveform diagrams are respectively of the noise reduction results of the method for the noisy frequency under the conditions of signal-to-noise ratio of 5dB, 10dB and 15dB. From the result waveforms, it can be seen that in the case of Example 3, the method can identify the main human voice audio and filter out other steady-state noise and other small-energy human voice noises. The human voice interference suppression ability is good. After the actual filtering, the auditory perception can obviously feel that the background noise intensity has been reduced. While the noise is reduced, the distortion of the voice part is small, which shows that the method has good speech recognition ability.

[0160] It should be noted that the present invention is not limited to the aforementioned specific embodiments and illustrated contents. After reading the contents of the specification, those skilled in the art may make various modifications, equivalent replacements or combined extensions without departing from the core ideas and technical solutions of the present invention, all of which should fall within the scope of protection of the present invention.

[0161] The embodiments shown in this specification and the drawings are for illustrative purposes only and should not be construed as limiting the scope of protection of the present invention.

Claims

1. A two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering, characterized in that: The following steps are involved: S1: Preprocess the received noisy speech signal, divide each frame of the speech signal into frequency bands after preprocessing, calculate the variance of each frame of the speech signal after frequency band division, perform logarithmic transformation on the variance of each frequency band to obtain the variance result of each frequency band, and judge whether the current frame is a speech frame or a noise frame based on the variance result of each frequency band and the set threshold; If the current frame is a noise frame, a static smoothing strategy is used to estimate the noise power spectrum of the noise frame; If the current frame is a speech frame, a dynamic smoothing strategy is used to estimate the speech power spectrum of the speech frame; Every Frame, Adaptive over-subtraction factor at each frequency point Dynamic adjustment based on signal-to-noise ratio: ； ； in , is an adjustable parameter, , , respectively controlling the steepness of the attenuation curve and the position of the attenuation threshold, Indicates the The first The noise power spectrum estimate of the frequency point, Indicates the The first The original signal power spectrum of the frequency point; A preliminary enhanced speech signal is obtained by calculating the adaptive over-subtraction factor and the enhanced spectrum subtraction formula; S2: Preprocess the initially enhanced speech signal, estimate the model parameters, and establish a model for Kalman filtering to obtain the denoised speech signal, specifically including: The initially enhanced speech signal is preprocessed to obtain a framed speech signal, and the speech signal model parameters are estimated for the framed speech signal using a linear predictive coding method; The process noise covariance parameters are estimated by taking the residual of the initially enhanced speech signal and the estimated speech signal model parameters, and the observation noise covariance parameters are estimated by the spectral flatness; A state-space model is constructed based on the speech signal model parameters, process noise covariance parameters and observation noise covariance parameters. Kalman filtering is performed on the state-space model to remove residual noise and obtain a denoised speech signal.

2. The two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering according to claim 1 is characterized in that: In steps S1 and S2, the preprocessing specifically includes: framing, windowing, pre-emphasis and short-time Fourier transform.

3. The two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering according to claim 1 is characterized in that: In step S1, the frequency band division adopts Bark frequency band division.

4. The two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering according to claim 1 is characterized in that: In step S1, the current frame is judged as a speech frame or a noise frame based on the comparison of the variance results of each frequency band with a set threshold, which specifically includes: When the variance result of the frequency band is greater than or equal to the threshold, the current frame is judged to be a speech frame; When the variance result of the frequency band is less than the threshold, the current frame is determined to be a noise frame.

5. The two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering according to claim 4 is characterized in that: In step S1, the threshold is set to -10~0dB.

6. A device for implementing the two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering according to any one of claims 1 to 5, characterized in that: include: Enhanced spectrum subtraction processing module, used to pre-process the noisy speech signal, estimate the noise power spectrum and speech power spectrum, and perform spectrum subtraction processing with an adaptive over-subtraction factor to obtain a preliminarily enhanced speech signal; The Kalman filter processing module is used to pre-process the initially enhanced speech signal, estimate model parameters, and establish a model for Kalman filter processing to obtain a noise-reduced speech signal.

7. A cochlear implant speech processing device comprising a microphone, a memory, and a processor, characterized in that: The microphone is used to collect voice signals in the environment, the memory is used to store programs, and the processor is used to execute the two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering as described in any one of claims 1 to 5.

8. The cochlear implant speech processing device according to claim 7, characterized in that: Also includes: A human brain electrode patch is connected to the processor, and the processor sends the noise-reduced voice signal to the human brain electrode patch.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the two-stage speech noise reduction method based on enhanced spectral subtraction and Kalman filtering according to any one of claims 1 to 5 is implemented.

Images