Speech denoising method and device based on mel subband cooperative period determination and double noise bounded fusion

Abstract

The application discloses a speech denoising method and device based on Mel sub-band cooperative period determination and double-noise bounded fusion. Full-bandwidth time-frequency analysis is performed on input audio; multiple Mel sub-band energy sequences are formed from input time domain samples, cross-sub-band cooperative period scores are formed, periodic speech state is determined and a fundamental frequency is determined; a harmonic protection mask is generated according to the fundamental frequency; frequency point differentiation update is performed, protection frequency points are frozen or limited slow update, and non-protection frequency points continue to track environmental noise; spectral differences between a candidate noise model and a working noise model are calculated on the non-protection frequency points, and continuous environmental changes are determined in combination with change frequency point coverage ratio and continuous frame number satisfaction; the candidate noise model is bounded and fused into the working noise model; spectral gain is calculated according to the updated working noise model, and noise reduction processing is performed on the complex frequency spectrum, and denoised audio is output through inverse transformation and overlap addition. Full-bandwidth, low-distortion and embedded real-time deployment suitable speech denoising is realized.

Description

Technical Field

[0001] This invention relates to the field of speech denoising technology, and in particular to a speech denoising method and apparatus based on Mel subband cooperative period determination and bounded fusion of dual noise. Background Technology

[0002] Mel subbanding refers to dividing the entire frequency band into multiple non-uniform subbands according to the Mel frequency scale, with finer subbands for low frequencies and coarser subbands for high frequencies, to simulate the human ear's "sensitive to low frequencies and insensitive to high frequencies" auditory characteristics. Mel subbanding is a non-uniform division, which differs from the traditional uniform subbanding (SB) approach; the latter divides the frequency band with equal width, potentially resulting in too many frequency units in the active low-frequency bands of speech, affecting overall performance. Single-channel frequency domain speech denoising typically includes framing, windowing, time-frequency transformation, noise power spectrum estimation, prior / posterior signal-to-noise ratio calculation, spectral gain generation, inverse transform, and overlap addition. Noise estimation determines the suppression benchmark for each frequency point; overestimation will attenuate the speech spectrum, while underestimation will result in residual noise.

[0003] Existing speech denoising technologies have the following shortcomings in terms of periodic speech preservation and environmental noise tracking: 1. Existing noise update methods struggle to address two opposing problems simultaneously. If the noise model continues to be updated according to uniform rules during periodic speech, stable vowels, long vowels, fundamental tones, and harmonics are easily treated as stable noise and written into the noise model, resulting in speech being swallowed and speech energy being excessively attenuated. If updates cease across the entire frequency range after periodic speech is detected, environmental changes such as air conditioning, wind noise, and road noise cannot be tracked, leading to residual new noise or continued excessive suppression after environmental noise decreases. Existing methods based on VAD (Voice Activity Detection) or SPP (Speech Presence Probability) cannot track environmental noise changes while preserving periodic speech.

[0004] 2. Existing methods lack fine-grained control. Using only the current frame's VAD, speech presence probability, or periodic binary results to control noise updates is prone to erroneous switching during brief drops in period determination, speech start / end boundaries, consonant transients, and abrupt changes in background noise. Current frame-by-frame unified control also fails to adequately distinguish between speech harmonic frequencies and dominant background frequencies, and lacks joint constraints on the duration of environmental changes, frequency band coverage, and model fusion rate.

[0005] 3. Existing methods face challenges in embedded deployment. When targeting 48kHz embedded processors, it is necessary to balance full-bandwidth real-time processing, rapid establishment of initial noise values ​​during startup, gain time-frequency smoothing, low-level fixed-point accuracy, and inverse transform overflow safety. Directly superimposing complex speech probability models, neural networks, or frequency-point nonlinear operations increases storage and computation per frame, which is detrimental to deployment on resource-constrained devices. Summary of the Invention

[0006] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a speech denoising method and apparatus based on Mel subband cooperative periodicity determination and bounded fusion of dual noise. The method generates a frequency-selective harmonic protection mask through Mel subband cooperative periodicity determination, and uses this mask to jointly control the differentiated update and bounded fusion of the frequency division points of the working noise model and the candidate noise model. While protecting the periodic speech pitch and harmonics from being absorbed by the noise model, the method continuously tracks changes in environmental noise, achieving full-bandwidth, low-distortion speech denoising suitable for embedded real-time deployment.

[0007] On one hand, embodiments of the present invention provide a speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise, including: Perform full-bandwidth time-frequency analysis on the input audio to obtain the complex spectrum and power spectrum at each frequency point; Multiple Mel subband energy sequences are formed from the input time-domain samples. Normalized time autocorrelation is performed on each Mel subband. The normalized time autocorrelation results of each Mel subband are summarized under the same candidate hysteresis to form a cross-subband cooperative periodic score. Based on the periodic score, cross-subband cooperativeness, fundamental frequency cross-frame continuity, and entry or exit hysteresis, the periodic speech state is determined and the fundamental frequency is identified. A frequency-selective harmonic protection mask covering the fundamental frequency, integer multiples of harmonics, and adjacent frequency points is generated based on the fundamental frequency. Simultaneously, a working noise model for spectral gain calculation and a candidate noise model for sensing environmental changes are maintained. During periodic speech, the working noise model and the candidate noise model are updated with frequency-division differentiation based on the protection mask, so that the protection frequency points are frozen or restricted to slow updates, while the non-protection frequency points continue to track environmental noise. The spectral difference between the candidate noise model and the working noise model is calculated at non-protected frequency points, and the continuous environmental changes are determined by combining the coverage ratio of the changing frequency points and the number of consecutively satisfied frames. When continuous environmental changes occur, the candidate noise model is boundedly fused into the working noise model according to the frequency division point fusion coefficient determined by the protection mask. The spectral gain is calculated based on the updated working noise model, and the complex spectrum is denoised. The denoised frequency is then output after inverse transformation and overlapping addition.

[0008] According to some embodiments of the present invention, the process of forming multiple Mel subband energy sequences from input time-domain samples includes: The input time-domain samples are filtered by a preset fixed-coefficient Mel subband filter bank to obtain the output signals of each Mel subband. The squared energy of each Mel subband output signal is calculated and grouped downsampled to obtain the energy sequence of each Mel subband.

[0009] According to some embodiments of the present invention, before the step of performing normalized time autocorrelation on each of the Mel subbands, the method further includes: Logarithmic compression and mean-reduction processing are performed on the energy sequences of each Mel subband.

[0010] According to some embodiments of the present invention, the formation of cross-subband cooperative periodicity scores includes: For each candidate lag, calculate the normalized time autocorrelation value of each Mel subband at the candidate lag; The median of the normalized time autocorrelation values ​​of each Mel subband is taken as the base score, and a coherence factor reflecting the consistency of support for the same lag in each Mel subband is introduced. The weighted sum of the basic score and the synergy factor is used as the periodic score of the candidate lag.

[0011] According to some embodiments of the present invention, the frequency-selective harmonic protection mask is a continuous weighted mask, and its protection bandwidth is adaptively adjusted according to the period confidence, wherein the period confidence is jointly determined by the period score, the cross-subband synergy, and the spectral peak confidence.

[0012] According to some embodiments of the present invention, the frequency-selective harmonic protection mask is a continuous weighted mask, and its protection bandwidth is adaptively adjusted according to the period confidence, wherein the period confidence is jointly determined by the period score, the cross-subband synergy, and the spectral peak confidence.

[0013] According to some embodiments of the present invention, calculating the spectral difference between the candidate noise model and the working noise model at the non-guard frequency point includes: The candidate noise model is continuously updated by the non-protected frequency during periodic speech, and the candidate background noise value at the protected frequency is supplemented by neighborhood statistics or frequency interpolation. The logarithmic power spectrum difference between the candidate noise model and the working noise model is calculated on the non-protected frequency point set, and the median of the logarithmic power spectrum difference is taken as the robust variation amplitude D(t). The coverage ratio R(t) of the frequency change points is the percentage of the logarithmic power spectrum difference that exceeds a predetermined threshold τ.d The proportion of the number of frequency points to the total number of frequency points in the set of non-protected frequency points; When the robust change amplitude D(t) and the change frequency coverage ratio R(t) simultaneously satisfy their respective thresholds τ D and τ R At that time, the cumulative number of consecutively satisfied frames is c(t); When the number of consecutive frames c(t) reaches the predetermined continuous frame count threshold L c When this condition is met, the determination of continuous environmental change is established.

[0014] According to some embodiments of the present invention, the frequency division point fusion coefficient determined according to the protection mask includes: The fusion coefficient λ of the protected frequency point harm The fusion coefficient λ is less than that of the non-guard frequency. env Furthermore, the fusion coefficient changes continuously with the weight value of the protection mask, that is: λ[k,t]=λ harm ·W[k,t]+λ env ·(1-W[k,t]), In the formula, λ[k,t] is the fusion coefficient of the k-th frequency point in the t-th frame; λ harm To protect the frequency fusion rate; W[k,t] is the protection mask weight for the k-th frequency point in the t-th frame; λ env This refers to the non-protected frequency fusion rate.

[0015] According to some embodiments of the present invention, after the step of boundedly fusing the candidate noise model into the working noise model, the method further includes: The maximum rise Δ per frame at each frequency point is applied to the fused working noise model. up and maximum decrease Δ down The limitation, and the final updated value, are: N w,new [k,t]=min(N w,old [k,t]+Δ up , max(N w,old [k,t]-Δ down Ñ w,new [k,t])), In the formula, N w,new [k,t] represents the new values ​​of the working noise model after bounded fusion; N w,old [k,t] represents the working noise model values ​​before fusion; w,new [k,t]) represents the fusion intermediate value; Δ up Δ is the maximum increase per frame per frequency point. down This represents the maximum drop in a single frame per frequency point.

[0016] On the other hand, embodiments of the present invention provide a speech denoising device based on Mel subband cooperative periodicity determination and bounded fusion of dual noise, for implementing the above-mentioned speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise. The system includes: An audio input module is used to receive input audio. The time-frequency analysis module is used to perform full-bandwidth time-frequency analysis on the input audio to obtain the complex spectrum and power spectrum at each frequency point; The Mel subband period detection module is used to form multiple Mel subband energy sequences from input time-domain samples, perform normalized time autocorrelation on each Mel subband, and summarize them under the same candidate hysteresis to form a cross-subband cooperative period score; determine the periodic speech state and determine the fundamental frequency based on the period score, cross-subband cooperativeness, fundamental frequency cross-frame continuity, and entry or exit hysteresis, and generate a frequency-selective harmonic protection mask based on the fundamental frequency; The candidate noise tracking module is used to maintain a candidate noise model for sensing environmental changes; The environmental change detection module is used to calculate the spectral difference between the candidate noise model and the working noise model at non-protected frequency points, and to determine the continuous environmental change by combining the coverage ratio of the changed frequency points and the number of consecutively satisfied frames. The model fusion module is used to maintain the working noise model for spectral gain calculation, update the frequency division point of the working noise model and the candidate noise model according to the protection mask during periodic speech, and fuse the candidate noise model into the working noise model in a bounded manner according to the frequency division point fusion coefficient determined by the protection mask when continuous environmental changes are established. The gain generation and smoothing module is used to calculate the spectral gain based on the updated working noise model and to perform noise reduction processing on the complex spectrum; The time-domain synthesis and output module is used to perform inverse transformation and superposition addition on the denoised complex spectrum to output the denoised frequency.

[0017] The embodiments of the present invention have at least the following beneficial effects: To address the technical contradiction of simultaneously protecting periodic speech and tracking environmental noise, a cross-subband collaborative periodic score is generated by performing normalized time autocorrelation on each Mel subband and summing the results under the same candidate time delay. This score, combined with dual thresholds for periodicity and collaborative degree, fundamental frequency cross-frame continuity, and entry / exit hysteresis determination, generates a frequency-selective harmonic protection mask. The same protection mask is used to jointly control the differentiated updates of the frequency division points of the working noise model and the candidate noise model, freezing or restricting the slow updates of the fundamental frequency, harmonics, and their adjacent frequencies, while non-protected frequencies continue to track background noise. When the variation amplitude, frequency coverage ratio, and duration between the candidate model and the working model meet the conditions, the candidate model is fused into the working model in a bounded manner according to the frequency division point fusion coefficient determined by the protection weight and the single-frame increase / decrease limit. Thus, during periodic speech, stable vowels, long tones, fundamental frequencies, harmonics, and speech tails are avoided from being absorbed by the noise model, while environmental changes such as air conditioning, wind noise, and road noise are continuously tracked. Meanwhile, by using cross-frame hysteresis, gradual mask attenuation, robust spectrum differences at unprotected frequency points, and joint determination of coverage ratio and number of consecutive frames, it suppresses erroneous switching caused by brief drops in period determination, speech boundaries, and consonant transients. Through single-path 1024-point full bandwidth analysis, fixed-coefficient Mel subband filtering, three-point group downsampling, lookup tables, static buffering, fixed-point format, and maximum safety compensation shift before inverse transformation, it significantly reduces storage and computation per frame while ensuring noise reduction performance, making it suitable for real-time deployment on resource-constrained embedded processors.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a speech denoising method based on Mel subband cooperative period determination and bounded fusion of dual noise, according to an embodiment of the present invention. Figure 2 This is a functional block diagram of a speech noise reduction device based on Mel subband cooperative period determination and dual noise bounded fusion according to an embodiment of the present invention. Figure 3 This is a flowchart of the period detection and harmonic protection mask generation process of the speech denoising method based on Mel subband cooperative period determination and dual noise bounded fusion according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the dual-noise model update and environmental change tracking of the speech denoising method based on Mel subband cooperative period determination and dual-noise bounded fusion according to an embodiment of the present invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0022] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," "exceeding," etc. are understood to exclude the stated number, and "above," "below," "within," etc. are understood to include the stated number. If "first," "second," etc. are used in the description, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the number of indicated technical features or the order of the indicated technical features.

[0023] In the description of this invention, unless otherwise explicitly defined, the terms "setting", "installing", "connecting" and "linking" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0024] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Please see Figure 1 This embodiment provides a speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise, mainly including steps S101~S107: S101. Perform full-bandwidth time-frequency analysis on the input audio to obtain the complex spectrum and power spectrum at each frequency point.

[0026] S102. Multiple Mel subband energy sequences are formed from the input time-domain samples. Normalized time autocorrelation is performed on each Mel subband. The normalized time autocorrelation results of each Mel subband are summarized under the same candidate hysteresis to form a cross-subband collaborative periodic score. Based on the periodic score, cross-subband collaborative degree, fundamental frequency cross-frame continuity, and entry or exit hysteresis, the periodic speech state is determined and the fundamental frequency is determined.

[0027] S103. Generate a frequency-selective harmonic protection mask that covers the fundamental frequency, integer multiples of harmonics, and adjacent frequency points based on the fundamental frequency.

[0028] S104. Simultaneously maintain the working noise model for spectral gain calculation and the candidate noise model for sensing environmental changes. During periodic speech, perform frequency point differentiation updates on the working noise model and the candidate noise model according to the protection mask, so that the protection frequency points are frozen or restricted to slow updates, while the non-protection frequency points continue to track environmental noise.

[0029] S105. Calculate the spectral difference between the candidate noise model and the working noise model at the non-protected frequency points, and determine the continuous environmental changes by combining the coverage ratio of the changing frequency points and the number of consecutively satisfied frames.

[0030] S106. When the continuous environmental change is established, the candidate noise model is boundedly fused into the working noise model according to the frequency division point fusion coefficient determined by the protection mask.

[0031] S107. Calculate the spectral gain based on the updated working noise model, and perform noise reduction processing on the complex spectrum. Output the noise-reduced frequency after inverse transformation and overlapping addition.

[0032] Please see Figure 2 This embodiment also provides a speech denoising device based on Mel subband cooperative periodicity determination and bounded fusion of two noises, used to implement the above-mentioned speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of two noises. The device includes: The audio input module 100 is used to receive input audio and acquire mono PCM data with a fixed frame length.

[0033] The time-frequency analysis module 200 is used to perform full-bandwidth time-frequency analysis on the input audio, save the previous frame input, concatenate it with the current frame, add a window, and then perform a real-number FFT to obtain the complex spectrum and power spectrum of each frequency point.

[0034] Mel subband period detection module 300 is used to form multiple Mel subband energy sequences from input time-domain samples, perform normalized time autocorrelation on each Mel subband, and summarize them under the same candidate hysteresis to form a cross-subband cooperative period score. Based on the period score, cross-subband cooperativeness, fundamental frequency cross-frame continuity, and entry / exit hysteresis, the periodic speech state is determined and the fundamental frequency is determined. Based on the fundamental frequency, a frequency-selective harmonic protection mask covering the fundamental tone, integer multiples of harmonics and their adjacent frequency points is generated.

[0035] The candidate noise tracking module 400 is used to maintain a candidate noise model for sensing environmental changes. During periodic speech, the candidate model is continuously updated by the non-protected frequency point, and the candidate background noise value at the protected frequency point is supplemented by neighborhood statistics or frequency interpolation.

[0036] The environmental change detection module 500 is used to calculate the spectral difference between the candidate noise model and the working noise model at the non-protected frequency points, and to determine the continuous environmental change by combining the coverage ratio of the changed frequency points and the number of consecutively satisfied frames.

[0037] The model fusion module 600 is used to maintain the working noise model for spectral gain calculation. During periodic speech, it performs frequency division point differential updates on the working noise model and candidate noise model according to the protection mask, so that the protection frequency points are frozen or restricted to slow updates, and the non-protection frequency points continue to track environmental noise. When continuous environmental changes are established, the candidate noise model is fused into the working noise model in a bounded manner according to the frequency division point fusion coefficient determined by the protection mask.

[0038] The gain generation and smoothing module 700 is used to calculate the spectral gain based on the updated working noise model, calculate the Wiener-type gain using decision-guided prior signal-to-noise ratio, apply a lower limit of gain and overdrive mapping, and then perform time smoothing and frequency smoothing before outputting the result.

[0039] The time-domain synthesis and output module 800 is used to perform inverse transformation and superposition addition on the denoised complex spectrum to output the denoised frequency.

[0040] The device also includes a strategy configuration module, which provides detection thresholds, noise update coefficients, model fusion rate, gain lower limit, and overdrive parameters to various related modules. Harmonic masks, candidate noise models, and environmental change determination jointly control the frequency division point update of the working noise model; the working noise model provides the noise power spectrum to the gain generation module; and the strategy configuration module provides update thresholds, fusion rate, gain lower limit, overdrive parameters, and mapping tables.

[0041] The following is a detailed description of the pipeline micro-corrosion leakage magnetic flux enhanced identification method and system provided in the embodiments of the present invention: 1. Receive mono audio and perform full-bandwidth time-frequency analysis Receive mono audio with a fixed frame length, for example, the sampling rate, frame length, analysis length, and number of frequency points are F. s =48000Hz, H=512, N=1024, and K=N / 2+1=513. The previous frame input is saved, concatenated with the current frame, and multiplied by a sinusoidal window of length N to obtain an analysis block. An N-point real-valued Fast Fourier Transform (FFT) is performed on the analysis block to obtain the complex spectrum and power spectrum. The power spectrum is used for updating the full-bandwidth main denoising chain and the working noise model. The periodic control chain obtains the sub-band energy sequence from the input time-domain samples. Both share the same frame boundary on the input side but do not mix detection variables.

[0042] Perform full-bandwidth time-frequency analysis on the input audio: Receive the current frame of a mono PCM with a sampling rate of 48kHz and a frame length of 512 points. Read the previous frame and concatenate it with the current frame in chronological order to form a 1024-point analysis block, then multiply by a sine analysis window. Perform a 1024-point real-number fast Fourier transform to obtain 513 non-redundant complex frequency points and calculate the power of each frequency point.

[0043] A three-layer collaborative processing is employed: The main denoising chain takes the full-bandwidth complex spectrum and power spectrum composed of the current frame and historical frames as input, calculates the frequency-by-frequency spectral gain based on the working noise model, and continuously outputs it. The periodic control chain takes the input time-domain samples as input, and outputs the periodic score, fundamental frequency trajectory, and frequency-selective harmonic protection mask through fixed-coefficient Mel subband filter banks, squared energy, three-point group downsampling, logarithmic compression, normalized time autocorrelation of each Mel subband, and consistency score. The model tracking chain updates the candidate noise model at non-protected frequencies, uses the robust spectral difference between the candidate model and the working model, the frequency coverage ratio, and the number of consecutive frames to determine environmental changes, and then fuses it into the working noise model according to the protection weight, frequency division rate, and single-frame amplitude limiting. The periodic control chain and the model tracking chain only change the update method of the noise parameters, while the main denoising chain continuously calculates the gain and outputs audio in all states.

[0044] 2. Periodic speech detection The periodic control branch filters the input time-domain samples using a preset group of M fixed-coefficient Mel subband filters. For example, M=24, covering 80–12000Hz.

[0045] The output of the m-th sub-band is: In the formula, x m [n] represents the filtered output of the m-th Mel sub-band; b m [q] represents the coefficients of the fixed subband filter, L b q is the subband FIR filter length, m is the Mel subband index, n is the time-domain sample index, q is the filter tap index, and x[n] is the input time-domain sample.

[0046] The subband energy sequence is obtained by grouping by square sum and three points: In the formula, E m [r] represents the energy sequence of the m-th Mel subband, D is the group downsampling factor, and D takes a value of 3; r is the detection sequence index, u is the index within the group; x m ²[] represents the subband output signal obtained after filtering the input signal by the m-th Mel subband filter.

[0047] Logarithmic compression and mean averaging across subbands are performed on the energies of each Mel subband to obtain: In the formula, E represents the energy of the m-th subband after logarithmic compression and mean subband subband sub-saturation. m [r] represents the energy sequence of the m-th Mel subband, ε is a positive decimal to prevent division by zero, j is the subband index when averaging, and r is the detection sequence index; E j [r] represents the energy value of the m-th Mel subband at the r-th detection time, which is obtained by downsampling the output signal of the subband through a sum of squares grouping.

[0048] The detection sampling rate is F det = F s / D. A candidate time delay set is formed for the 80-400Hz fundamental frequency range, and the candidate time delays are selected to satisfy F. det / 400 ≤ l ≤ F det Let l be an integer of / 80, where l is the candidate delay. Let the maximum candidate delay be l. max = F det / 80, to ensure that the maximum candidate delay has complete relevant samples, the detection sequence length R is... e Satisfy R e ≥ l max +1.

[0049] For each candidate time delay l, calculate the normalized time autocorrelation for each subband: In the formula, R m (l) represents the normalized temporal autocorrelation value of the m-th Mel subband at candidate time delay l. is Candidate delay, r is the detection sequence index, R l The set of valid indices for the summation. m [r] represents the normalized logarithmic energy value of the m-th Mel subband at the r-th detection time, after logarithmic compression and mean removal across subbands; m [rl] represents the energy sequence value of the same sub-band before a delay of l detection points, i.e., the value of the m-th sub-band at r. The normalized logarithmic energy at time l. Summation is performed only if both r and rl fall within the range of the valid detection sequence.

[0050] It should be noted that the aforementioned "cross-subband consistency" is not cross-correlation between different subbands, but rather the summing of the results after calculating the normalized autocorrelation of each Mel subband along the time axis, under the same candidate lag value l. The periodic score only improves when multiple subbands simultaneously support the same lag value. This mechanism utilizes the physical characteristic that the pitch period of speech exists simultaneously in multiple Mel subbands, effectively suppressing accidental correlations that only occur in a single subband.

[0051] Periodic scoring is defined as: In the formula, S(l) is the periodic score corresponding to the candidate time delay l. This represents the median of the normalized time autocorrelation value of each Mel subband at time delay l; η is The synergy weighting coefficient is set to 0.25; Q(l) represents the cross-subband synergy.

[0052] in, , representing the degree of coordination between different Mel subbands for the same period. In the formula, For the j-th Mel subband in time delay l l Normalized time autocorrelation value at [location].

[0053] Pick When S(l) ≥ T per And Q(l) ≥ Q min The current frame period candidates are formed at this time. Here, l* is the optimal candidate delay that maximizes the period score S(l). For example, T per =0.80, Q min =0.65.

[0054] Then f0 = F det / l* The fundamental frequency is obtained, and a protection mask is generated by combining the fundamental frequency continuity across frames, periodic scoring, and spectral peak confidence. Here, f0 is the fundamental frequency calculated from the optimal time delay; F det is the detection sampling rate; l* is the optimal candidate delay that maximizes the periodic score S(l).

[0055] Let the optimal fundamental frequency of the current frame be f0(t), the fundamental frequency of the candidate frame in the previous effective period be f0(t-1), and the relative fundamental frequency deviation be: δ f (t) = |f0(t) - f0(t-1)| / max(f0(t-1), ε); In the formula, δ f (t) represents the relative deviation of the current frame's base frequency from the base frequency of the previous valid frame; ε is a positive decimal to prevent division by zero; T f For the relative fundamental frequency deviation threshold, 0 <Tf <1, with a value of 0.15.

[0056] When δf(t) ≤ T f The fundamental frequency is assumed to be continuous across frames. If there is no historical valid fundamental frequency, the current fundamental frequency is used as the first frame of the candidate trajectory, and the cross-frame deviation is not calculated; the relative fundamental frequency deviation is calculated in subsequent frames.

[0057] Periodic scoring, cross-subband coherence, and cross-frame continuity conditions continuously satisfy L on After the frame, periodic protection begins; if any condition is not met consecutively, L... off Exit periodic protection after frame, where L on ≥1 and L off ≥1 are all integers, thus creating entry and exit lag.

[0058] The protection mask generation uses harmonic center frequency and confidence level adaptive bandwidth.

[0059] The center frequency of the h-th harmonic is f h = h·f0.

[0060] Let Δf = F s / N is the FFT frequency resolution, and the joint confidence of period, subband consistency, and spectral peak confidence is: In the formula, C is the joint confidence level of period, subband consistency, and spectral peak confidence; C p C represents the confidence level of the spectral peak. p ∈[0,1]; T per is the periodic scoring threshold, with a value of 0.80; S The period score is calculated at the optimal delay. The cross-subband coordination degree at the optimal delay.

[0061] The protected bandwidth is: B h = B0 + κ(1-C)Δf ; In the formula, B h denoted as the protection bandwidth for the h-th harmonic, B0 as the minimum protection bandwidth, κ as the confidence expansion coefficient, C as the joint confidence level, and Δf as the FFT frequency resolution.

[0062] Frequency-by-frequency protection weights are: In the formula, This indicates that the maximum value of h is taken for all harmonic orders; W[k] is the protection weight at the k-th frequency point, 0 ≤ W [ k ]≤1;B h f is the protection bandwidth for the h-th harmonic;k f is the frequency of the k-th frequency point; h is the center frequency of the h-th harmonic.

[0063] The closer W[k] is to 1, the more likely that the frequency point belongs to periodic speech. h=1,...,H h f represents the harmonic order not exceeding the Nyquist frequency. k = kF s / N represents the frequency of the k-th frequency point, where k=0,...,K-1.

[0064] Please see Figure 3 First, compare whether the periodic score and cross-subband coordination degree have reached the dual thresholds. Then, combine the fundamental frequency cross-frame continuity and the entry / exit hysteresis to determine whether the periodic protection state is established. If it is established, calculate the fundamental frequency f0=F using the optimal time delay. det / l*, and adaptively determine the harmonic protection bandwidth based on the spectral peak confidence, forming a frequency-by-frequency protection weight W[k] for the fundamental tone, integer multiple harmonics, and their neighborhoods. If not true, the history mask is not immediately cleared, but rather smoothly decayed frame by frame to cover the boundaries of transient drops, voiced tails, and periodic speech ends. For example, T per =0.80, Q min =0.65, T f =0.15, L on =3 frames, L off =5 frames, that is, the periodic protection state is entered after the conditions of periodic score not less than 0.80, coordination degree not less than 0.65, and relative fundamental frequency deviation not exceeding 15% are met for 3 consecutive frames; the periodic protection state is exited after any condition is not met for 5 consecutive frames.

[0065] 3. Periodic voice protection and environmental change tracking state machine Configure a state machine with five states: Start-up learning, normal update, periodic selective protection, environmental change tracking, and recovery fusion.

[0066] Let the operating noise power spectrum be N. w [k] is used for spectral gain calculation in the current frame; the candidate noise power spectrum is N. c [k] is used to sense background changes; the harmonic protection weight is W[k], 0≤W[k]≤1, where a value close to 1 indicates that the frequency point is more likely to belong to periodic speech.

[0067] (1) Initiation of learning phase For candidate noise models N respectively c [k] and operating noise model N w [k] Perform cumulative learning and impose a limit on the single-frame increase of the working model. For example, learn the frame threshold L. max=282 frames, corresponding to approximately 3.008 seconds (each frame is 512 / 48000 seconds long). Let L cnt This represents the number of valid learning frames accumulated up to the current frame. The initial value is 0, and it increments by 1 after each frame is completed and learning begins. Let N be the number of frames accumulated up to the current frame. floor [k]>0 represents the minimum noise power per frequency point preset according to the fixed-point quantization accuracy, denoted as Δ. up ≥0 represents the maximum single-frame increase allowed per frequency point in the working model.

[0068] First, calculate the candidate cumulative value using the current power P[k]: In the formula, This represents the cumulative estimate of the candidate noise model; This represents the number of valid learning frames accumulated before the current frame, initially set to 0. These are the old values ​​for the candidate noise model; The power of the k-th frequency point in the current frame; The minimum noise power preset for the k-th frequency point; L max The threshold for initiating the learning frame is set to 282 frames, approximately 3.008 seconds.

[0069] The candidate model is then updated as follows: N c,new [k] = max(N est [k], N floor [k]); In the formula, N c,new [k] represents the new value of the candidate noise model.

[0070] The working model is then updated using the new values ​​from the candidate model and the old values ​​from the working model: N w,new [k] = min(N w,old [k] + Δ up , max(N c,new [k], N floor [k])) ; In the formula, N w,new [k] represents the new value of the working noise model, N w,old [k] represents the old value of the working noise model; Δ up This represents the maximum per-frame increase allowed per frequency point in the operating model, approximately 3 dB.

[0071] Finally, the updated value is written back to the old value in the next frame. Thus, even if the initial N... w Even if [k] is zero or close to zero, a non-zero initial noise value can be established, while preventing the working model from being momentarily boosted by sudden speech bursts during the startup phase.

[0072] Please seeFigure 4 First, determine if the learning process has been completed. If the learning process has not been completed (the number of learning frames has not reached L), then... max Simultaneously, the working noise model and candidate noise models are accumulated, and the single-frame rise rate Δ of the working model is limited. up This prevents sudden increases in initial noise levels during the power-on phase. For example, L max =282 frames (approximately 3.008 seconds), Δ up This corresponds to a maximum increase of 3dB per frame per frequency point (power domain ratio of approximately 2.0 times).

[0073] (2) Normal update phase When no periodic speech is detected, the regular update value N is first calculated based on the relationship between the current power and the old value of the working model. u [k].

[0074] When P[k]>N w,old When [k], a slow ascending recursion is used: N u [k] = α up ·N w,old [k] + (1-α up )·P[k]; When P[k] ≤ N w,old When [k], a fast descent recursion is used: N u [k] = α down ·N w,old [k] + (1-α down )·P[k]; In the formula, N u [k] represents the regular update value; α up α is the slowly increasing recurrence relation coefficient. up The value is 0.999; α down The rapidly decreasing recurrence relation, α down =0.96.

[0075] The candidate model adopts the same directional correlation recursive structure, and the old values ​​of the working model are replaced with the old values ​​of the candidate model; the rise and fall coefficients of the candidate model can be the same as those of the working model, or they can be configured as independent coefficients that are more sensitive to environmental changes.

[0076] After the learning process is completed, the candidate noise model is continuously updated using non-harmonic frequencies. The candidate background values ​​of the protected frequency points are supplemented by neighborhood median, quantile statistics or frequency interpolation, so that the candidate model can sense environmental changes without absorbing long tone and harmonic energy.

[0077] (3) Periodic selective protection stage The full-frequency freeze is no longer used; the working model is updated as follows: N w,new [k] = W[k]·N w,old [k] + (1-W[k])·N u [k]; In the formula, N w,new [k] represents the new value of the working noise model, N w,old [k] represents the old value of the working noise model; N u [k] is the regular update value; W[k] is the protection mask weight. When W[k] is close to 1, the working model keeps the original value or is subject to restricted slow updates.

[0078] Therefore, harmonic frequencies and their adjacent frequencies remain unchanged or are updated only at a limited, slow rate, while non-harmonic frequencies continue to track background noise. The protection mask employs cross-frame hysteresis and gradual attenuation to cover periodic drops and speech tails.

[0079] (4) Continuous updating of candidate models Candidate Model N c [k] is still updated by non-harmonic frequencies during periodic speech, and the background noise at the protected frequency is estimated by neighborhood median, quantile statistics, or frequency interpolation to avoid writing harmonic energy into the candidate model. The candidate update can use a different time constant than the working model, making it more sensitive to environmental changes but not directly determining the suppression gain of the current frame.

[0080] (5) Environmental change monitoring Environmental change detection in the set of non-harmonic frequencies Ω nh (t) = {k∈{0,...,K-1}: W[k,t] ≤ T nh On}, calculate the logarithmic power difference between the candidate model and the working model, where Ω nh (t) represents the set of unprotected frequency points in the t-th frame; W[k,t] represents the protection mask weight of the k-th frequency point in the t-th frame; T nh The threshold for determining non-protected frequency points, T nh ∈[0,1].

[0081] definition: d k (t) = log((N c [k,t]+ε) / (N w [k,t]+ε)), D(t) = median{k∈Ω nh (t)} |d k (t)| , R(t) = |{k∈Ω nh (t): |dk (t)| ≥ τ d}| / |Ω nh (t)|; In the formula, d k (t) represents the logarithmic power difference between the candidate model and the working model at the k-th frequency, and is N. c [k,t] represents the candidate noise model and the working noise model; D(t) represents the robust variation amplitude at the unprotected frequency point; R(t) represents the coverage ratio of the variation frequency point.

[0082] The number of consecutive frames is updated according to the following formula: when D(t) ≥ τ D And R(t) ≥ τ R When c(t) = c(t-1)+1, then c(t) = c(t-1)+1; otherwise c(t) = 0.

[0083] Where c(t) is the number of frames that continuously satisfy the environmental change conditions; τ d For single-frequency difference threshold, τ d >0, with a value of 0.5; τ D For a robust change range threshold, τ D >0, with a value of 0.3; τ R The frequency coverage ratio threshold is 0≤τ R ≤1, with a value of 0.4.

[0084] When c(t) ≥ L c When the background noise changes continuously, it is determined that L is the noise level. c L is an integer continuous frame count threshold. c ≥1, with a value of 10. This persistence and frequency band coverage combined determination can exclude false triggering caused by single consonants, transients, and residual harmonics.

[0085] For the set of nonharmonic frequencies Ω nh (t) Calculate the logarithmic spectrum difference between the candidate model and the working model, and characterize persistence by robust variation amplitude, coverage ratio of variation frequencies, and number of consecutively satisfied frames. When the effective non-harmonic frequencies are less than K... min When this happens, the change determination for this frame is set to invalid. K min K is the minimum number of effective unprotected frequency points. min ≥1. For example, K min =50、τ d =0.5 (when the power ratio is calculated using the natural logarithm, this value corresponds to a power domain variation of approximately 2.2 dB), τ D =0.3、τ R =0.4, L c =10 frames. That is, provided that the number of unprotected frequency points is not less than 50, the median difference in log power spectrum between the candidate model and the working model is not less than 0.3, and the difference exceeds τ.d The frequency ratio of (i.e., 0.5) is not less than 40%, and the environment is determined to have changed continuously after the above conditions are met for 10 consecutive frames.

[0086] (6) Environmental change tracking and bounded integration After determining that there is an environmental change, the system enters the environmental change tracking state. First, the frequency division point fusion coefficient is defined: λ[k,t] = λ harm ·W[k,t] + λ env ·(1-W[k,t]); In the formula, λ[k,t] is the fusion coefficient of the k-th frequency point in the t-th frame; λ harm To protect the frequency fusion rate, a value of 0.01 is used; W[k,t] is the protection mask weight for the k-th frequency point in the t-th frame; λ env The unprotected frequency fusion rate is set to 0.20; satisfying 0 ≤ λ. harm <λ env ≤1.

[0087] Then, the candidate noise model is written into the working noise model: Ñ w,new [k,t] = (1-λ[k,t])·N w,old [k,t] + λ[k,t]·N c [k,t]; In the formula, Ñ w,new [k,t] represents the intermediate values ​​of the fusion of candidate models into the working model.

[0088] To limit single-frame mutations, the final updated value is: N w,new [k,t]=min(N w,old [k,t]+Δ up , max(N w,old [k,t]-Δ down Ñ w,new [k,t])) ; Where, N w,new [k,t] represents the new values ​​of the working noise model after bounded fusion; Δ up The maximum rise time per frame at each frequency point is approximately 3 dB; the maximum fall time per frame at each frequency point is approximately 6 dB. 0 ≤ λ harm <λ env ≤1, Δ up ≥0 and Δ down ≥0 represent the maximum rise and fall of a single frame per frequency point, respectively. Since the frequency points where W[k,t] is close to 1 are guard frequencies, their fusion rate is close to λ. harm The fusion rate of unprotected frequency points is close to λ. env`min()` and `max()` are amplitude limiting operators: `max()` applies a lower limit to prevent excessively rapid drops; `min()` applies an upper limit to prevent excessively rapid increases. Together, they ensure that the change in the noise model per frame does not exceed Δup and Δdown, thus avoiding abrupt changes in noise floor and gain.

[0089] After the periodic speech ends, the working noise model enters the recovery fusion state, decreasing W[k,t] and increasing λ[k,t] frame by frame, so that all frequency points smoothly return to normal tracking, avoiding noise floor and gain jump caused by state switching.

[0090] When the continuous change holds true, candidate models are boundedly fused into the working model according to harmonic protection weights, frequency division point fusion rates, and per-frame increment / decrement limits. For example, λ harm =0.01、λ env =0.20, meaning the fusion rate for guarded frequencies is only 1% / frame, while the fusion rate for unguarded frequencies is 20% / frame. Maximum rise in a single frame Δ up and maximum decrease Δ down The limits are set at 3dB and 6dB per frequency point (power domain), respectively. If the continuous change does not hold, the periodic speech state is further assessed.

[0091] When periodic speech is present, harmonic frequencies are frozen or restricted to slow updates, while non-harmonic frequencies are updated normally. When periodic speech is absent, a full-frequency slow rise (α) is performed. up =0.999), rapid decline (α) down Asymmetric update (=0.96).

[0092] Then, a continuous, constrained working noise model that can track environmental changes is obtained. When the periodic speech ends or the protection state is released, the protection weight is reduced frame by frame, so that the update rate of the harmonic frequency points gradually returns to the normal value, avoiding sudden changes in noise floor and spectral gain caused by state switching.

[0093] 4. Decision-guided spectral gain and dual-domain smoothing Calculate the posterior signal-to-noise ratio at the k-th frequency point: γ[k] = P[k] / (N w [k] + ε); In the formula, γ[k] is the posterior signal-to-noise ratio at the k-th frequency point; P[k] is the power spectrum of the current frame; N w [k] represents the operating noise model.

[0094] And let the instantaneous prior terms be: ξ inst [k] = max(γ[k] - 1, 0); In the formula, ξ inst [k] represents the instantaneous prior signal-to-noise ratio term.

[0095] Utilizing the smoothed gain G from the previous frame prev [k] and the posterior signal-to-noise ratio γ of the previous frame prev [k] Forms the prior signal-to-noise ratio for decision guidance: ξ[k] = β·G prev [k]²·γ prev [k] + (1-β)·ξ inst [k]; In the formula, ξ[k] is the decision-guided prior signal-to-noise ratio; β is the decision-guided smoothing coefficient, with a value of 0.98; G prev [k] represents the smoothing gain of the previous frame; γ prev This is the a posteriori signal-to-noise ratio of the previous frame.

[0096] The base gain adopts a Wiener-type relationship: G w [k] = ξ[k] / (1+ξ[k]); In the formula, G w [k] represents the Wiener-type base gain.

[0097] First, apply the lower gain limit G according to the suppression strategy. min Then perform driver mapping: G p [k] = max(G min G w [k] q ) ; In the formula, G p [k] represents the gain after overdrive mapping; q represents the overdrive parameter. q >0, with a value of 1.45; G min The lower limit of gain is 0 ≤ G min ≤1, with a value of 0.10.

[0098] To avoid calling the exponentiation function for every frame and frequency point, a mapping table of 257 entries is pre-generated during initialization or policy update. At runtime, this is adjusted according to G. w Linear interpolation is performed at position [k]. Where 0 ≤ G min ≤1, and q>0.

[0099] Time smoothing uses: G t [k] = 0.75·G prev [k] + 0.25·G p [k]; In the formula, G t [k] represents the gain after time smoothing.

[0100] Subsequently, the internal frequency points are smoothed using the five points with the highest center weight on the frequency axis: G s [k]=0.10·G t [k-2] + 0.20·G t [k-1] + 0.40·G t [k] + 0.20·G t [k+1] +0.10·G t [k+2]; In the formula, G s [k] represents the final gain after frequency smoothing.

[0101] The boundary frequencies are respectively treated with either a three-point or two-point normalized kernel. The final gain is limited to [G]. min The range of [1] is multiplied by the complex spectrum. Temporal smoothing reduces frame-by-frame pumping, and five-point frequency smoothing suppresses musical noise caused by isolated spectral peaks.

[0102] The posterior signal-to-noise ratio γ[k] = P[k] / (N) is calculated using the current power spectrum and the operating noise model. w [k]+ε), and combined with the smoothing gain G from the previous frame. prev [k] Formation of decision-guided prior signal-to-noise ratio ξ[k]=β·G prev [k]²·γ prev [k]+(1-β)·max(γ[k]-1,0). For example, β=0.98.

[0103] Generate a Wiener-type base gain Gw[k]=ξ[k] / (1+ξ[k]) with a lower gain limit, and perform continuous overdrive mapping Gp[k]=max(G) through a lookup table according to the suppression strategy. min G w [k] q ).

[0104] First, perform cross-frame temporal smoothing G. t [k]=0.75·G prev [k]+0.25·G p [k], then use the five frequency kernels with the highest center weights for adjacent frequency smoothing to reduce music noise caused by gain pumping and isolated spectral peaks: G s [k]=0.10·G t [k-2]+0.20·G t [k-1]+0.40·G t [k]+0.20·G t [k+1]+0.10·G t [k+2]; In the formula, Gs [k] is the final gain after frequency smoothing, G t [k] represents the gain after time smoothing.

[0105] 5. Continuous mapping of suppression intensity The suppression policy value ranges from 0 to 255. When policy = 0, the output gain is set to 1, which is equivalent to bypass noise reduction gain, but input analysis, state update, and noise model maintenance are still maintained. When 1 ≤ policy ≤ 255, piecewise interpolation is performed based on the lower limit of gain and overdrive parameters defined by multiple anchor points.

[0106] First, calculate the normalized position u between adjacent anchor points, then use s=u 2 (3-2u) is used for smooth transition to obtain G respectively. min And q. Where s is the smooth transition value; u is the normalized position between adjacent anchor points. For example, policy=170 corresponds to G min =0.10, q=1.45.

[0107] Continuous mapping allows the product to adjust the suppression intensity without interrupting the noise history, avoiding parameter jumps caused by discrete gear switching.

[0108] 6. Inverse Transformation and Overlapping Addition The final gain is multiplied by both the real and imaginary parts of the corresponding frequency point, a 1024-point inverse real FFT is performed, multiplied by the same sine synthesis window as the analysis window, and then overlapped with the latter half of the previous analysis block and added. The first 512 points are output as the 16-bit PCM for this frame, and the last 512 points are carried over to the next frame. The sine analysis window and the synthesis window satisfy a square complementarity under a 50% overlap condition, thus achieving continuous reconstruction when the gain is 1.

[0109] The smoothing gain is applied to both the real and imaginary parts of the complex spectrum. The numerical boundaries of the DC frequency, Nyquist frequency, and all complex frequencies are checked. The maximum safe compensation shift without overflow is selected, and then the inverse real FFT is performed.

[0110] Multiply by the synthesis window and add it with 50% overlap to the latter half of the previous analysis block. Perform rounding, saturation, and 16-bit quantization to output a 512-point PCM. Save the current input frame, previous frame gain, overlap buffer, working noise model, candidate noise model, and guard mask state, and then proceed to the next frame loop.

[0111] 7. Fixed-point accuracy and overflow control For FFT implementations that scale the inverse transform input, a downward search begins from the predetermined maximum compensation bit depth before the inverse transform. Candidate left-shift bits must simultaneously satisfy the following conditions: the real and imaginary parts of the DC frequency (index 0), the Nyquist frequency (index N / 2), and the recombined pairs derived from the real-number FFT symmetry (i.e., the sum and difference of X[k] and X[Nk]) after left shifting are all within the safe range of a 32-bit signed integer. The maximum number of bits satisfying these conditions is selected as the compensation shift, and the remaining compensation amount after the inverse transform is recorded to ensure that subsequent accumulation does not overflow. This frame-by-frame selection of the maximum safe compensation allows the low-level spectrum to fully utilize the dynamic range of parameter Q31, while explicit boundary checks prevent inverse transform input overflow.

[0112] Fixed-point implementations can choose between a general-purpose KISS FFT backend or a processor-specific real-number FFT backend. Backend switching does not change the logic relationships of period protection, noise state machine, signal-to-noise ratio estimation, gain generation, and overlap addition. During runtime, statically allocated window tables, spectrum arrays, noise arrays, gain arrays, period detection buffers, and overlap addition buffers are used, eliminating the need to allocate or release heap memory per frame.

[0113] For example, window coefficients, spectral gain, and nonlinear mapping tables are stored in Q15 fixed-point format; time-domain windowed data and complex spectra are stored in Q31 fixed-point format; power is stored in unsigned Q30 fixed-point format; and signal-to-noise ratio is stored in unsigned Q16 fixed-point format. Squaring, correlation, and accumulation are performed using 64-bit intermediate variables. All multiplications undergo rounding and saturation processing. Wiener-type gain can be calculated directly using 64-bit division, or rewritten as 1 minus the reciprocal term when the processor has 32-bit hardware division, thus reducing the division bit width.

[0114] The Mel period detection branch prequantizes the coefficients of each subband filter into a Q15 fixed-point format and stores them statically. The squared energy of the subband is accumulated using 64 bits and then normalized to an unsigned Q30 fixed-point format. The reciprocal and square root in the logarithmic compression and autocorrelation scoring are approximated using lookup tables and fixed-point approximations. The detection branch downsamples at D=3, outputting only the period score and fundamental frequency trajectory, without changing the 513 full-bandwidth frequency points of the main processing chain. This increases the period protection decision capability while avoiding runtime floating-point filter banks and dynamic memory allocation.

[0115] To verify the technical effects of the present invention, the following verifications were conducted: 1. Voice protection and environmental tracking verification The input includes sustained vowels and voiced tails, with background noise superimposed within the spoken segment, exhibiting step-increase, slow-increase, and sudden-decrease patterns. By comparing three schemes—conventional continuous update, full-frequency freezing, and the selective dual-model tracking of this invention—the noise model drift near the fundamental frequency and several harmonics, environmental tracking delay in non-harmonic frequency bands, output residual noise, and speech distortion are statistically analyzed.

[0116] Verification results: The noise model drift of the present invention is significantly smaller than that of the conventional continuous update scheme, the environmental tracking delay is significantly smaller than that of the full-frequency freezing scheme, and the output speech has lower distortion and less noise residue.

[0117] 2. Music noise verification By comparing the outputs with no smoothing, time smoothing only, and time plus five-point frequency smoothing, the number of isolated spectral peaks, spectral gain variance, and subjective music noise score were statistically analyzed.

[0118] Verification results: The music noise score of the time-plus-five-point frequency smoothing scheme is significantly better than other schemes.

[0119] 3. Fixed-point consistency verification Compare the signal-to-noise ratio, maximum absolute error, and saturation order of floating-point and fixed-point outputs, and cover silent, full-sine, impulse, and random noise inputs.

[0120] Verification results: The signal-to-noise ratio difference between fixed-point output and floating-point output is within 1dB, with no abnormal saturation.

[0121] 4. Resource Verification Record the processing cycle, peak static storage, and number of heap allocations for each 512-point frame.

[0122] Verification results: The single-frame processing cycle meets the real-time requirements (approximately 10.67ms per frame at a 48kHz sampling rate), with no heap memory allocation and peak static storage suitable for the target embedded processor.

[0123] The embodiments of the present invention have the following beneficial effects: 1. To address the technical challenge of simultaneously satisfying periodic speech protection and environmental noise tracking, this paper employs Mel subband cooperative periodicity determination, dual thresholds for periodicity scoring and cooperativeness, fundamental frequency cross-frame continuity and entry / exit hysteresis determination, and frequency-selective harmonic protection masks. At protected frequencies, the working noise model is frozen or restricted to slow updates, while candidate noise models continue to be updated at non-protected frequencies. When environmental noise undergoes continuous changes, candidate models are then fused into the working model in a bounded manner according to protection weights. This simultaneously prevents stable vowels, long vowels, fundamental tones, harmonics, and speech tails from being absorbed by the noise model, while ensuring that background changes can still be tracked during periodic speech.

[0124] 2. To ensure that the above core innovations continue to work even when the period determination is unstable, the cross-Mel subband period score, fundamental frequency cross-frame continuity, cross-frame hysteresis and gradual attenuation of harmonic protection mask, robust spectral differences on non-harmonic frequencies, frequency coverage ratio exceeding the threshold, number of consecutively satisfied frames, and frequency division fusion rate are combined for noise parameter update control, thereby suppressing erroneous switching caused by brief drops in period determination, speech boundaries, consonant transients, and sudden changes in background noise.

[0125] 3. To enable the above core innovations to be implemented on resource-constrained processors, a single 1024-point main analysis path is used to cover the full bandwidth from 0 to 24 kHz. The periodic control branch uses low-order fixed-coefficient Mel subband filtering, three-point grouping downsampling, and fixed-point autocorrelation. The main processing chain uses decision-guided gain, time smoothing, and five-point frequency smoothing to reduce gain jitter and music noise. The embedded implementation is completed using lookup tables, static working buffers, fixed-point format, and maximum safety compensation shift before inverse transformation.

[0126] 4. Compared with existing noise update methods based on VAD or SPP, this invention exhibits stronger robustness in non-stationary noise environments. Existing methods either freeze updates across the entire frequency range during periodic speech (unable to track environmental changes) or continue updating across the entire frequency range (speech is absorbed by the noise model), while this invention achieves a balance between the two through frequency-selective protection and a dual-model architecture.

[0127] 5. Compared with neural network-based speech enhancement methods, it does not require pre-trained models and a large amount of computing resources, making it suitable for real-time deployment on resource-constrained devices such as embedded DSPs and microcontrollers.

[0128] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise, characterized in that, include: Perform full-bandwidth time-frequency analysis on the input audio to obtain the complex spectrum and power spectrum at each frequency point; Multiple Mel subband energy sequences are formed from the input time-domain samples. Normalized time autocorrelation is performed on each Mel subband. The normalized time autocorrelation results of each Mel subband are summarized under the same candidate hysteresis to form a cross-subband cooperative periodic score. Based on the periodic score, cross-subband cooperativeness, fundamental frequency cross-frame continuity, and entry or exit hysteresis, the periodic speech state is determined and the fundamental frequency is identified. A frequency-selective harmonic protection mask covering the fundamental frequency, integer multiples of harmonics, and adjacent frequency points is generated based on the fundamental frequency. Simultaneously, a working noise model for spectral gain calculation and a candidate noise model for sensing environmental changes are maintained. During periodic speech, the working noise model and the candidate noise model are updated with frequency-division differentiation based on the protection mask, so that the protection frequency points are frozen or restricted to slow updates, while the non-protection frequency points continue to track environmental noise. The spectral difference between the candidate noise model and the working noise model is calculated at non-protected frequency points, and the continuous environmental changes are determined by combining the coverage ratio of the changing frequency points and the number of consecutively satisfied frames. When continuous environmental changes occur, the candidate noise model is boundedly fused into the working noise model according to the frequency division point fusion coefficient determined by the protection mask. The spectral gain is calculated based on the updated working noise model, and the complex spectrum is denoised. The denoised frequency is then output after inverse transformation and overlapping addition.

2. The speech denoising method based on Mel subband collaborative periodicity determination and bounded fusion of dual noise as described in claim 1, characterized in that, The process of forming multiple Mel subband energy sequences from input time-domain samples includes: The input time-domain samples are filtered by a preset fixed-coefficient Mel subband filter bank to obtain the output signals of each Mel subband. The squared energy of each Mel subband output signal is calculated and grouped downsampled to obtain the energy sequence of each Mel subband.

3. The speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise as described in claim 2, characterized in that, Before the step of performing normalized time autocorrelation on each of the Mel subbands, the method further includes: Logarithmic compression and mean-reduction processing are performed on the energy sequences of each Mel subband.

4. The speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise as described in claim 3, characterized in that, The formation of cross-subband coordinated periodic scores includes: For each candidate lag, calculate the normalized time autocorrelation value of each Mel subband at the candidate lag; The median of the normalized time autocorrelation values ​​of each Mel subband is taken as the base score, and a coherence factor reflecting the consistency of support for the same lag in each Mel subband is introduced. The weighted sum of the basic score and the synergy factor is used as the periodic score of the candidate lag.

5. The speech denoising method based on Mel subband collaborative periodicity determination and bounded fusion of dual noise as described in claim 1, characterized in that, The frequency-selective harmonic protection mask is a continuous weighted mask, and its protection bandwidth is adaptively adjusted according to the period confidence. The period confidence is jointly determined by the period score, the cross-subband synergy, and the spectral peak confidence.

6. The speech denoising method based on Mel subband collaborative periodicity determination and bounded fusion of dual noise as described in claim 1, characterized in that, The determination of the periodic voice state adopts entry hysteresis and exit hysteresis: When the periodic score, the cross-subband coordination degree, and the fundamental frequency cross-frame continuity condition all continuously satisfy the first predetermined frame number L on Then, it enters periodic protection mode; When any condition fails to meet the second predetermined frame number L consecutively off Afterwards, exit the cycle protection state, where L on and L off All are greater than or equal to 1 and are all integers.

7. The speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise according to claim 1, characterized in that, The step of calculating the spectral difference between the candidate noise model and the working noise model at the non-protected frequency point includes: The candidate noise model is continuously updated by the non-protected frequency during periodic speech, and the candidate background noise value at the protected frequency is supplemented by neighborhood statistics or frequency interpolation. The logarithmic power spectrum difference between the candidate noise model and the working noise model is calculated on the non-protected frequency point set, and the median of the logarithmic power spectrum difference is taken as the robust variation amplitude D(t). The coverage ratio R(t) of the frequency change points is the percentage of the logarithmic power spectrum difference that exceeds a predetermined threshold τ. d The proportion of the number of frequency points to the total number of frequency points in the set of non-protected frequency points; When the robust change amplitude D(t) and the change frequency coverage ratio R(t) simultaneously satisfy their respective thresholds τ D and τ R At that time, the cumulative number of consecutively satisfied frames is c(t); When the number of consecutive frames c(t) reaches the predetermined continuous frame count threshold L c When this condition is met, the determination of continuous environmental change is established.

8. The speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise according to claim 1, characterized in that, The frequency division point fusion coefficients determined according to the protection mask include: The fusion coefficient λ of the protected frequency point harm The fusion coefficient λ is less than that of the non-guard frequency. env Furthermore, the fusion coefficient changes continuously with the weight value of the protection mask, that is: λ[k,t]=λ harm ·W[k,t]+λ env ·(1-W[k,t]), In the formula, λ[k,t] is the fusion coefficient of the k-th frequency point in the t-th frame; λ harm To protect the frequency fusion rate; W[k,t] is the protection mask weight for the k-th frequency point in the t-th frame; λ env This refers to the non-protected frequency fusion rate.

9. The speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise according to claim 8, characterized in that, After the step of boundedly fusing the candidate noise model into the working noise model, the method further includes: The maximum rise Δ per frame at each frequency point is applied to the fused working noise model. up and maximum decrease Δ down The limitation, the final updated value is: N w,new [k,t]=min(N w,old [k,t]+Δ up , max(N w,old [k,t]-Δ down , Ñ w,new [k,t])), In the formula, N w,new [k,t] represents the new values ​​of the working noise model after bounded fusion; N w,old [k,t] represents the working noise model values ​​before fusion; w,new [k,t]) represents the fusion intermediate value; Δ up Δ is the maximum increase per frame per frequency point. down This represents the maximum drop in a single frame per frequency point.

10. A speech denoising device based on Mel subband cooperative periodicity determination and bounded fusion of dual noise, characterized in that, The speech denoising method based on Mel subband cooperative periodicity determination and bounded fusion of dual noise as described in any one of claims 1 to 9 includes: An audio input module is used to receive input audio. The time-frequency analysis module is used to perform full-bandwidth time-frequency analysis on the input audio to obtain the complex spectrum and power spectrum at each frequency point; The Mel subband period detection module is used to form multiple Mel subband energy sequences from input time-domain samples, perform normalized time autocorrelation on each Mel subband, and summarize them under the same candidate hysteresis to form a cross-subband cooperative period score; determine the periodic speech state and determine the fundamental frequency based on the period score, cross-subband cooperativeness, fundamental frequency cross-frame continuity, and entry or exit hysteresis, and generate a frequency-selective harmonic protection mask based on the fundamental frequency; The candidate noise tracking module is used to maintain a candidate noise model for sensing environmental changes; The environmental change detection module is used to calculate the spectral difference between the candidate noise model and the working noise model at non-protected frequency points, and to determine the continuous environmental change by combining the coverage ratio of the changed frequency points and the number of consecutively satisfied frames. The model fusion module is used to maintain the working noise model for spectral gain calculation, update the frequency division point of the working noise model and the candidate noise model according to the protection mask during periodic speech, and fuse the candidate noise model into the working noise model in a bounded manner according to the frequency division point fusion coefficient determined by the protection mask when continuous environmental changes are established. The gain generation and smoothing module is used to calculate the spectral gain based on the updated working noise model and to perform noise reduction processing on the complex spectrum; The time-domain synthesis and output module is used to perform inverse transformation and superposition addition on the denoised complex spectrum to output the denoised frequency.

Images