A signal denoising method, device, equipment, medium and product

Abstract

The application provides a signal noise reduction method, device, equipment, medium and product, and relates to the technical field of voice signal processing. The method generates a first noise estimation result focusing on music noise suppression and a second noise estimation result focusing on noise mutation tracking through a dual-path noise estimation unit, and dynamically regulates the update speed of the first noise estimation result based on the second noise estimation result, so that the smoothing process can be strengthened to deeply suppress music noise in a stationary noise scene, and the update speed of the first noise estimation result is accelerated to realize rapid tracking when noise mutates. In addition, the dual-path cooperative architecture breaks through the dependence of single-path noise estimation on fixed parameters, can adapt to the dynamic changes of different noise states, avoids noise residues or tracking delays caused by scene switching, finally realizes the improvement of the naturalness and clarity of the noise-reduced voice in all scenes, and optimizes the subjective listening experience of the user.

Description

Technical Field

[0001] This application relates to the field of speech signal processing technology, and in particular to a signal noise reduction method, apparatus, device, medium and product. Background Technology

[0002] Currently, in single-channel speech denoising technology, speech enhancement algorithms based on short-time spectrum estimation are the main technical means to achieve noise suppression and are widely used in the speech processing modules of various terminals. These algorithms obtain the noise power spectrum through noise estimation and then combine it with the characteristics of the noisy speech signal to separate noise from clean speech, thereby achieving a noise reduction effect. The performance of noise estimation directly determines the overall noise reduction effect and the subjective listening experience of the speech.

[0003] To improve noise estimation performance, various noise estimation algorithms have been proposed in existing technologies, such as the minimum value controlled recursive averaging algorithm, the improved minimum value controlled averaging algorithm, and the unbiased minimum mean square error algorithm. However, these algorithms struggle to simultaneously achieve both effective noise suppression and noise mutation tracking, impacting subjective listening experience. Furthermore, existing single-channel noise reduction schemes often employ a single-path noise estimation architecture, which is ill-suited to dynamically changing real-world noise scenarios, resulting in poor subjective listening experience of the denoised speech. Summary of the Invention

[0004] This application provides a signal noise reduction method, apparatus, device, medium, and product to solve the problems in the prior art.

[0005] In a first aspect, this application provides a signal noise reduction method, comprising:

[0006] Acquire noisy speech domain signals;

[0007] The noisy speech domain signal is input to a dual-channel noise estimation unit for noise processing to obtain a first noise estimation result and a second noise estimation result; wherein, the first noise estimation result is an estimation result used to suppress music noise, and the second noise estimation result is an estimation result used to improve the ability to track noise abrupt changes;

[0008] Based on the first noise estimation result and the second noise estimation result, the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result is adjusted.

[0009] The noise power spectrum is determined based on the first noise estimation result after update speed adjustment;

[0010] Based on the noisy speech domain signal and the noise power spectrum, the denoised time-domain output signal is determined.

[0011] In one possible design, adjusting the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result based on the first noise estimation result and the second noise estimation result includes:

[0012] Extract the first speech presence probability from the first noise estimation result, and extract the second speech presence probability from the second noise estimation result;

[0013] Based on the probability of the first speech occurrence and the probability of the second speech occurrence, determine whether the first noise estimation result is in a near-stagnant state;

[0014] Based on the judgment result of the near-stagnation state, the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result is adjusted.

[0015] In one possible design, extracting a first speech presence probability from the first noise estimation result and extracting a second speech presence probability from the second noise estimation result includes:

[0016] The real-time speech absence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio are extracted from the first noise estimation result. Based on the real-time speech absence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio, a mapping calculation is performed to obtain the first speech presence probability.

[0017] The second real-time posterior signal-to-noise ratio is extracted from the second noise estimation result. Based on the second real-time posterior signal-to-noise ratio, the first preset value and the second preset value are mapped and calculated to obtain the probability of the second speech presence. The first preset value is a fixed probability value of the speech not being present, and the second preset value is a fixed prior signal-to-noise ratio value.

[0018] In one possible design, determining whether the first noise estimation result is in a near-stagnant state based on the first speech presence probability and the second speech presence probability includes:

[0019] Calculate the whole-frame distribution distance between the probability of the first speech occurrence and the probability of the second speech occurrence;

[0020] Extract the feature parameters corresponding to the whole frame distribution distance; wherein, the feature parameters include at least one of mean, variance, sub-band mean difference, and Itakula distance;

[0021] Based on the characteristic parameters, determine whether the first noise estimation result is in a near-stagnant state.

[0022] In one possible design, adjusting the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result based on the judgment result of the near-stagnation state includes:

[0023] If the first noise estimation result is determined to be in a near-stagnant state, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit is increased;

[0024] If it is determined that the first noise estimation result is not in a near-stagnant state, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit remains unchanged.

[0025] In one possible design, determining the denoised time-domain output signal based on the noisy speech domain signal and the noise power spectrum includes:

[0026] The posterior signal-to-noise ratio is calculated based on the amplitude spectrum of the noisy speech domain signal and the noise power spectrum.

[0027] The prior signal-to-noise ratio is determined based on the posterior signal-to-noise ratio;

[0028] The Wiener gain coefficient is calculated based on the prior signal-to-noise ratio.

[0029] Based on the Wiener gain coefficient and the noisy speech domain signal, a clean speech domain signal is obtained;

[0030] Perform an inverse Fourier transform on the pure speech audio domain signal to obtain a time-domain speech signal;

[0031] The time-domain speech signal is subjected to frame synthesis processing to obtain a denoised time-domain output signal.

[0032] Secondly, this application provides a signal noise reduction device, comprising:

[0033] The acquisition module is used to acquire noisy speech audio domain signals;

[0034] The noise processing module is used to input the noisy speech domain signal to the dual-channel noise estimation unit for noise processing to obtain a first noise estimation result and a second noise estimation result; wherein, the first noise estimation result is an estimation result used to suppress music noise, and the second noise estimation result is an estimation result used to improve the ability to track noise abrupt changes;

[0035] The control module is used to control the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result based on the first noise estimation result and the second noise estimation result.

[0036] The noise power spectrum determination module is used to determine the noise power spectrum based on the first noise estimation result after the update rate adjustment.

[0037] The time-domain output signal determination module is used to determine the denoised time-domain output signal based on the noisy speech domain signal and the noise power spectrum.

[0038] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0039] The memory stores computer-executed instructions;

[0040] The processor executes computer execution instructions stored in the memory to implement the method as described in any of the first aspects.

[0041] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0042] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the first aspects.

[0043] This application provides a signal noise reduction method, apparatus, device, medium, and product. The method generates a first noise estimation result focusing on music noise suppression and a second noise estimation result focusing on noise mutation tracking through dual-channel noise estimation units. Based on the second noise estimation result, the update speed of the first noise estimation result is dynamically adjusted. This allows for enhanced smoothing processing to deeply suppress music noise in stable noise scenarios, while accelerating the update speed of the first noise estimation result to achieve rapid tracking during noise mutations. This effectively solves the performance conflict between music noise suppression and noise mutation tracking in traditional algorithms. Furthermore, the dual-channel collaborative architecture breaks through the dependence of single-channel noise estimation on fixed parameters, enabling it to adapt to dynamic changes in different noise states and avoid noise residue or tracking delay caused by scene switching. Ultimately, it improves the naturalness and clarity of the denoised speech in all scenarios, optimizing the user's subjective listening experience. Attached Figure Description

[0044] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0045] Figure 1 This application provides an example of a signal noise reduction method according to one embodiment of the present application, and includes an application scenario diagram.

[0046] Figure 2 A schematic flowchart illustrating a signal noise reduction method provided in an embodiment of this application;

[0047] Figure 3 A schematic flowchart of a signal noise reduction method provided in another embodiment of this application;

[0048] Figure 4 This is a schematic diagram of the structure of a signal noise reduction device provided in an embodiment of this application;

[0049] Figure 5 This is a structural example diagram of an electronic device provided in an embodiment of this application.

[0050] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0051] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0052] To clearly understand the technical solution of this application, the solutions of the prior art will be described in detail first.

[0053] In single-channel speech denoising technology, speech enhancement algorithms based on short-time spectrum estimation are the main technical means to achieve noise suppression and are widely used in the speech processing modules of various terminals. These algorithms obtain the noise power spectrum through noise estimation and then combine it with the characteristics of the noisy speech signal to separate noise from clean speech, thereby achieving a noise reduction effect. The performance of noise estimation directly determines the overall noise reduction effect and the subjective listening experience of the speech.

[0054] To improve noise estimation performance, various noise estimation algorithms have been proposed in existing technologies, such as the minimum value controlled recursive averaging algorithm, the improved minimum value controlled averaging algorithm, and the unbiased minimum mean square error algorithm. However, these algorithms struggle to simultaneously achieve both effective music noise suppression and noise mutation tracking capabilities, resulting in a significant performance trade-off. Specifically, algorithms that prioritize music noise suppression typically require improving the smoothness of noise estimation, which significantly increases noise tracking latency. In stable noise scenarios where noise levels suddenly increase, noise estimation can easily become near-stagnant. In such cases, the system may misclassify sudden pure noise as noisy speech, further reducing the update speed of noise estimation and creating a vicious cycle of underestimation, leading to a further increase in noise tracking latency. On the other hand, algorithms that prioritize improving noise mutation tracking capabilities and reducing tracking latency, due to insufficient smoothing of noise estimation, can lead to a significant decrease in music noise suppression capabilities, resulting in noticeable music noise residue in the denoised speech and affecting subjective listening experience.

[0055] Furthermore, existing single-channel noise reduction solutions mostly employ a single-path noise estimation architecture, attempting to balance the needs of music noise suppression and noise mutation tracking by adjusting only a single parameter such as update window length and recursive averaging time. However, noise conditions in real-world applications are constantly and dynamically changing. Fixed parameters optimized for a single scenario cannot cover the suppression requirements of different noise scenarios. Music noise residue is prone to appear during periods of stable noise, while tracking issues are likely to occur during periods of sudden noise changes, making it difficult to guarantee stable noise reduction performance across all scenarios.

[0056] In summary, existing single-channel speech denoising technology cannot effectively suppress music noise while ensuring rapid tracking of noise abrupt changes. Furthermore, the architecture of single-channel noise estimation is difficult to adapt to dynamically changing real-world noise scenarios, resulting in poor subjective listening experience of the denoised speech.

[0057] Figure 1 An application scenario diagram corresponding to a signal denoising method provided in an embodiment of this application is shown, such as... Figure 1 As shown, the application scenario provided in this embodiment includes: a noisy voice acquisition device 10, a signal noise reduction processing device 11, and a voice output terminal 12. The noisy voice acquisition device 10 and the signal noise reduction processing device 11 interact through an audio data transmission link, and the signal noise reduction processing device 11 and the voice output terminal 12 are connected through an audio decoding link. The noisy voice acquisition device 10 can be a mobile phone microphone, a Bluetooth headset pickup module, a smart speaker pickup unit, a vehicle-mounted voice acquisition device, etc.; the voice output terminal 12 can be a mobile phone earpiece, a vehicle-mounted voice broadcaster, or a smart speaker sound unit.

[0058] First, the noisy speech acquisition device 10 acquires noisy speech signals from the environment through a single channel. Then, it sends the noisy speech signals to the signal denoising processing device 11. The signal denoising processing device 11 first converts the noisy speech signals into noisy audio domain signals to prepare for subsequent noise estimation. Then, it inputs the frequency domain signal into the dual-channel noise estimation unit to simultaneously obtain two types of estimation results: a first noise estimation result focusing on music noise suppression and a second noise estimation result focusing on rapid tracking of noise abrupt changes. Next, the signal denoising processing device 11 dynamically adjusts the update speed of the first noise estimation result based on the characteristic differences between the two types of results, balancing smoothness and tracking performance. Then, it calculates the noise power spectrum based on the adjusted first noise estimation result. Finally, the signal denoising processing device 11 combines the noisy audio domain signal and the noise power spectrum to complete noise separation, restores the processed frequency domain signal to the denoised time domain output signal, and transmits it to the speech output terminal 12 for playback or recognition.

[0059] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0060] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0061] Figure 2 This is a flowchart illustrating a signal denoising method according to an embodiment of this application, as shown below. Figure 2 As shown, the execution subject of this embodiment is a signal noise reduction device. This device can be implemented by a computer program, or by a medium storing the relevant computer program, such as a USB flash drive and / or optical disc; alternatively, it can be implemented by a physical device integrating or installing the relevant computer program, such as a chip or electronic device. The electronic device may be a computer or a server, etc. The signal noise reduction method provided in this embodiment includes the following steps:

[0062] S201. Obtain the noisy speech domain signal.

[0063] Optionally, the noisy speech time-domain signal can be acquired through the microphone of the terminal device. This noisy speech time-domain signal is a superposition signal of clean speech signal and environmental noise signal.

[0064] Optionally, the acquired noisy speech time-domain signal is preprocessed. The preprocessing steps may include framing, windowing, and Fast Fourier Transform (FFT). Framing involves dividing the continuous time-domain signal into frames of fixed length (typically 10-30ms), with approximately 50% overlap between adjacent frames to avoid the impact of abrupt changes in the signal between frames. Windowing multiplies each frame signal by a window function such as a Hanning window or a Hamming window to reduce spectral leakage. Fast Fourier Transform converts each frame of the time-domain signal into a frequency-domain signal, obtaining the noisy speech audio-visual domain signal.

[0065] It should be noted that speech signals and noise signals have different distribution characteristics in the frequency domain. For example, speech signals are mainly concentrated in the 100Hz-4000Hz frequency band, while noise signals have a wider frequency band distribution. Therefore, it is easier to separate noise from speech in the frequency domain, which lays the foundation for subsequent noise estimation and noise reduction processing.

[0066] S202. The noisy speech domain signal is input to the dual-channel noise estimation unit for noise processing to obtain a first noise estimation result and a second noise estimation result; wherein, the first noise estimation result is an estimation result used to suppress music noise, and the second noise estimation result is an estimation result used to improve the ability to track noise mutations.

[0067] This step replaces the single-path noise estimation architecture in the existing technology by setting up dual parallel noise estimation branches, decoupling the two conflicting performance goals of music noise suppression and noise mutation tracking, which are then implemented by two independent noise estimation branches.

[0068] The dual-path noise estimation unit includes a first noise estimation branch and a second noise estimation branch. The two branches receive noisy speech signals in parallel, execute independent noise estimation algorithms, and output noise estimation results with different focuses.

[0069] Specifically, the first noise estimation result is used to suppress musical noise. The corresponding first noise estimation branch adopts a noise estimation algorithm with strong smoothing constraints, which can be achieved by setting a large smoothing factor and a long recursive averaging time. Strong smoothing constraints can keep the first noise estimation result stable and with small fluctuations, thereby effectively suppressing musical noise in the denoised speech and ensuring the cleanliness and naturalness of the subjective listening experience of the speech.

[0070] Specifically, the second noise estimation result is used to improve the ability to track noise abrupt changes. Its corresponding second noise estimation branch employs a noise estimation algorithm with weak smoothing constraints, which can be achieved by setting a small smoothing factor and a short recursive averaging time. The weak smoothing constraint enables the second noise estimation result to quickly respond to changes in noise intensity in noisy speech signals, especially in scenarios with sudden noise changes. It can quickly track increases in noise intensity and promptly output estimation results that closely approximate the true noise level.

[0071] It should be noted that the first and second noise estimation branches can be improved based on existing noise estimation algorithms. Different performance emphases can be achieved simply by adjusting parameters such as the smoothing factor and recursive averaging duration, without designing entirely new noise estimation algorithms, thus reducing the design complexity and implementation difficulty. Existing noise estimation algorithms can include MCRA (Minima Controlled Recursive Averaging), IMCRA (Improved Minima Controlled Recursive Averaging), and UMMSE (Unbiased Minimum Mean Squared Error).

[0072] S203. Based on the first noise estimation result and the second noise estimation result, adjust the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit.

[0073] Specifically, this step adaptively adjusts the update speed of the first noise estimation branch through mutual feedback between the two noise estimation results, so that the first noise estimation result can maintain strong smoothness to suppress music noise, and can be updated quickly to track noise changes when noise changes abruptly.

[0074] Optionally, when the ambient noise is stable, the noise intensity in the noisy speech signal changes little. At this time, the deviation between the second noise estimation result (fast tracking branch) and the first noise estimation result (strong smoothing branch) is small, indicating that there are no sudden noise changes. In this case, the first noise estimation branch is adjusted to work at a slower update speed to maintain its strong smoothing characteristics and avoid noise estimation fluctuations caused by excessively fast updates, thereby effectively suppressing music noise.

[0075] Optionally, when environmental noise undergoes a sudden change (such as a sudden increase in noise intensity), the second noise estimation branch, due to weak smoothing constraints, will respond quickly to the noise change, and its output second noise estimation result will rise rapidly, increasing the deviation from the first noise estimation result. This deviation signal is the detection signal for the noise sudden change. At this time, based on the magnitude of the deviation between the two estimation results, the update speed of the first noise estimation branch is adaptively increased, allowing the first noise estimation result to quickly approach the true noise level and avoiding incomplete noise suppression due to tracking delay.

[0076] Furthermore, the update speed can be controlled by adjusting the smoothing factor of the first noise estimation branch. The larger the deviation, the smaller the smoothing factor, and the faster the update speed; the smaller the deviation, the larger the smoothing factor, and the slower the update speed, thus ensuring the accuracy and real-time nature of the control.

[0077] S204. The noise power spectrum is determined based on the first noise estimation result after update rate adjustment.

[0078] Specifically, after adaptive adjustment by S203, the first noise estimation result can maintain strong smoothness in stable noise scenarios, effectively suppressing musical noise; in scenarios with sudden noise changes, the first noise estimation result can also achieve rapid tracking of noise changes by increasing the update speed, avoiding underestimation of noise. Therefore, determining the final noise power spectrum using the first noise estimation result after update speed adjustment can balance the accuracy of noise suppression and the naturalness of speech sound.

[0079] It should be noted that the noise power spectrum is the main input parameter for subsequent speech enhancement processing, and its accuracy directly determines the noise reduction effect.

[0080] S205. Based on the noisy speech domain signal and the noise power spectrum, determine the time-domain output signal after noise reduction.

[0081] Optionally, a speech enhancement algorithm based on short-time spectral estimation is employed, combining the noisy speech audio domain signal and a determined noise power spectrum to suppress noise components in the noisy speech audio domain signal. The speech enhancement algorithm can be spectral subtraction, Wiener filtering, minimum mean square error spectral estimation, etc. Based on the noise power spectrum, the speech enhancement algorithm calculates the distribution ratio of noise in the noisy speech audio domain signal, and then weights and corrects the noisy speech audio domain signal to remove noise components and retain clean speech components.

[0082] Specifically, after completing the frequency domain noise reduction process, the processed frequency domain signal is subjected to an inverse fast fourier transform (IFFT) to convert the frequency domain signal into a time domain signal. Subsequently, the time domain signal is frame-overlapped and added to complete the time domain reconstruction, and finally the noise-reduced time domain speech signal is output. This time domain speech signal has the characteristics of thorough noise suppression, no obvious music noise, and natural subjective listening experience.

[0083] This application provides a signal denoising method that generates a first noise estimation result focusing on music noise suppression and a second noise estimation result focusing on noise mutation tracking through dual-path noise estimation units. The update speed of the first noise estimation result is dynamically adjusted based on the second noise estimation result. This allows for enhanced smoothing processing to deeply suppress music noise in stable noise scenarios, while accelerating the update speed of the first noise estimation result to achieve rapid tracking during noise mutations. This effectively solves the performance conflict between music noise suppression and noise mutation tracking in traditional algorithms. Furthermore, the dual-path collaborative architecture overcomes the dependence of single-path noise estimation on fixed parameters, adapting to dynamic changes in different noise states and avoiding noise residue or tracking delays caused by scene switching. Ultimately, it improves the naturalness and clarity of the denoised speech across all scenarios, optimizing the user's subjective listening experience.

[0084] As an optional implementation, based on any of the above embodiments, the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result is adjusted according to the first noise estimation result and the second noise estimation result, including the following steps:

[0085] First, the probability of the presence of the first speech is extracted from the first noise estimation result, and the probability of the presence of the second speech is extracted from the second noise estimation result.

[0086] Specifically, the probability of speech presence refers to the probability that clean speech exists in the noisy speech domain signal of the current frame. Its value ranges from [0,1]. The closer the value is to 1, the greater the probability that clean speech exists in the current frame; the closer the value is to 0, the greater the probability that the current frame is pure noise. The extraction of the probability of speech presence is based on the first noise estimation result, the second noise estimation result, and the power spectrum characteristics of the noisy speech domain signal, combined with existing speech detection algorithms, such as energy-based speech detection and spectral entropy-based speech detection.

[0087] Optionally, the ratios of the power spectrum of the noisy speech signal to the first noise estimation result and the second noise estimation result are calculated respectively to obtain the first power spectrum ratio and the second power spectrum ratio. These ratios reflect the relative intensity of the speech component and the noise component in the noisy speech; a larger ratio indicates a higher proportion of the speech component and a greater probability of the presence of clean speech. Next, the first and second power spectrum ratios are normalized to eliminate amplitude differences under different frequency bands and noise intensities, ensuring the consistency and comparability of the extracted speech presence probabilities. Then, based on the normalized power spectrum ratios, the first speech presence probability (corresponding to the first noise estimation result) and the second speech presence probability (corresponding to the second noise estimation result) are calculated using a preset probability mapping function (such as the sigmoid function or a linear mapping function). The first speech presence probability has a more stable value and smaller fluctuations due to the strong smoothing characteristic of the first noise estimation result; the second speech presence probability has a faster response due to the weak smoothing characteristic of the second noise estimation result, and can quickly follow the switching between speech and noise.

[0088] Optionally, to extract the probability of the first speech occurrence from the first noise estimation result, it is necessary to first perform amplitude spectrum estimation on the noisy speech audio domain signal to obtain the power spectrum of the noisy speech, which lays the foundation for subsequent parameter extraction. The amplitude spectrum estimation formula is as follows:

[0089]

[0090] Where α is the smoothing coefficient, ranging from 0 to 1, with a fixed value typically chosen as 0.9. This represents a convolution operation, where b is the convolution kernel; based on the estimated amplitude spectrum... Combined with minimum noise tracking value Extracting the probability that real-time voice does not exist ,in,

[0091]

[0092] Alternatively, the probability of the absence of speech can be calculated using linear mapping, sigmoid mapping, or other methods.

[0093] It should be noted that the extraction of the probability of speech presence does not require additional complex computational logic. It can be directly derived based on existing speech detection algorithms and two-channel noise estimation results without increasing the algorithm's complexity.

[0094] Secondly, based on the probability of the first speech occurrence and the probability of the second speech occurrence, it is determined whether the first noise estimation result is in a near-stagnant state.

[0095] The near-stagnation state refers to a critical state in which the first noise estimation branch, due to strong smoothing constraints, has an extremely low update rate, and its output noise estimation result cannot respond to noise changes in a timely manner, about to fall into stagnation (update rate approaches 0). This is also the initial state of the vicious cycle of "noise underestimation" in the existing technology. If this near-stagnation state can be detected in time and control can be triggered, the first noise estimation result can be effectively prevented from falling into complete stagnation, thus avoiding the vicious cycle of underestimation.

[0096] Optionally, if within a consecutive preset time threshold, the difference in the probability of speech existence is continuously greater than the first preset threshold, and the probability of the first speech existence is continuously lower than the second preset threshold, it indicates that the current second noise estimation branch has detected noise change (or pure noise), while the first noise estimation branch, due to the strong smoothing constraint, has not responded in time, and its probability of speech existence is still at a low level (representing misjudgment as pure noise or close to pure noise). At this time, it can be determined that the first noise estimation result is in a near-stagnant state.

[0097] The preset time threshold can be 3 frames, corresponding to 60ms, to adapt to the short-term characteristics of the voice signal; the voice presence probability difference is the difference between the second voice presence probability and the first voice presence probability; the first preset threshold can be set to 0.15, the second preset threshold can be set to 0.05, and the first preset threshold is greater than the second preset threshold.

[0098] Optionally, if the probability difference of the speech is not continuously greater than the first preset threshold, or the probability of the first speech is not continuously lower than the second preset threshold, it indicates that the current first noise estimation branch can still respond normally to noise changes and is not in a state of imminent stagnation, so there is no need to trigger emergency control.

[0099] It should be noted that the probability of the first speech is stable and the response is slow, while the probability of the second speech is fast and can reflect noise changes in a timely manner. When the difference between the two is consistently large and the probability of the first speech is consistently low, it indicates that the first noise estimation branch has failed to keep up with the pace of noise changes, the update speed is close to stagnation, and it is unable to update the noise estimation results in a timely manner, which will soon lead to a vicious cycle of underestimation.

[0100] Finally, based on the judgment result of the near-stagnation state, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit is adjusted.

[0101] Optionally, based on the judgment result of the near-stagnation state, a differentiated control strategy can be adopted to ensure that the update is triggered quickly in the near-stagnation state and to avoid miscontrol in the non-stagnation state, so as to ensure that the first noise estimation branch achieves a precise balance between "strong smoothing to suppress music noise" and "fast tracking of noise mutation".

[0102] Optionally, if the first noise estimation result is determined to be in a near-stagnant state, emergency control is triggered to reduce the smoothing factor of the first noise estimation branch (e.g., from 0.98 to 0.80-0.85) to improve its update speed, so that the first noise estimation result can quickly approach the real noise level corresponding to the second noise estimation result and quickly get out of the near-stagnant state; at the same time, the weak smoothing constraint of the second noise estimation branch remains unchanged to continuously provide real-time reference for noise mutations.

[0103] Optionally, if it is determined that the first noise estimation result is not in a state of near stagnation, the original control logic of the first noise estimation branch is maintained, that is, the smoothing factor is adaptively adjusted according to the magnitude of the deviation between the two noise estimation results to maintain its strong smoothing characteristics and ensure the music noise suppression effect; at this time, there is no need to make emergency adjustments to avoid the first noise estimation result from fluctuating more due to over-adjustment, which would generate music noise.

[0104] Furthermore, after triggering emergency control, the changes in the probability of the first speech presence and the probability of the second speech presence are monitored in real time. When the difference in the probability of speech presence is less than the first preset threshold and the probability of the first speech presence rises back to above the second preset threshold, it indicates that the first noise estimation result has escaped the near-stagnation state, gradually restores its original smoothing factor, and returns to the normal control mode, taking into account both noise tracking and music noise suppression.

[0105] This application provides a signal denoising method that extracts the speech presence probability from dual-path noise estimation results and determines whether the first noise estimation result is in a near-stagnant state. It then dynamically adjusts the update speed of the first noise estimation result. This allows for timely acceleration of the first noise estimation result's update to quickly respond to sudden noise changes when noise tracking is detected to be about to stall due to excessive smoothing during music noise suppression. In non-stagnant states, the original smoothing characteristics are maintained to ensure effective music noise suppression. Thus, the adaptive adjustment mechanism effectively solves the problem of balancing smoothing processing and tracking response in traditional algorithms, improves the robustness of noise estimation in complex dynamic scenarios, and optimizes the speech quality and subjective listening experience after denoising.

[0106] As an optional implementation, based on any of the above embodiments, extracting the first speech presence probability from the first noise estimation result and the second speech presence probability from the second noise estimation result includes the following steps:

[0107] Specifically, the real-time speech non-existence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio are extracted from the first noise estimation result. Based on the real-time speech non-existence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio, a mapping calculation is performed to obtain the first speech existence probability.

[0108] The first noise estimation result includes the noise power spectrum features corresponding to the noisy speech domain signal of the current frame. The real-time speech non-existence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio are extracted from the noise power spectrum features.

[0109] The probability of real-time speech absence refers to the probability that the current frame is pure noise and there is no clean speech (value range [0,1]), which is complementary to the probability of speech presence. The real-time prior signal-to-noise ratio (SNR) is the ratio of the power spectrum of clean speech to the power spectrum of noise in the current frame, reflecting the inherent intensity relationship between speech and noise. The first real-time posterior SNR is the ratio of the power spectrum of noisy speech in the current frame to the first noise estimation result (strongly smoothed noise power spectrum), reflecting the actual proportion of speech and noise intensity in the current frame.

[0110] Optionally, the real-time prior signal-to-noise ratio is extracted from the first noise estimation result. First real-time posterior signal-to-noise ratio First, calculate the intermediate parameters of the probability of the first speech occurrence. :

[0111]

[0112] Then, based on the probability of the absence of real-time speech, the real-time prior signal-to-noise ratio, and intermediate parameters, a mapping calculation is performed to obtain the probability of the presence of the first speech. The specific mapping formula is as follows:

[0113]

[0114] Optionally, a preset probability mapping formula is used to fuse the probability of the absence of real-time speech, the real-time prior signal-to-noise ratio, and the first real-time posterior signal-to-noise ratio to obtain the probability of the presence of the first speech.

[0115] It should be noted that the extraction of the above feature parameters is based on the power spectrum characteristics of the first noise estimation result itself, without the need to collect other signals. The calculation logic is simple and can be adapted to the strong smoothness characteristics of the first noise estimation branch, ensuring the stability of the probability of the first speech and providing reliable support for the accurate judgment of the subsequent near-stagnation state.

[0116] Specifically, the second real-time posterior signal-to-noise ratio is extracted from the second noise estimation result, and a mapping calculation is performed based on the second real-time posterior signal-to-noise ratio, the first preset value, and the second preset value to obtain the probability of the second speech presence; wherein, the first preset value is a fixed probability value of the speech not being present, and the second preset value is a fixed prior signal-to-noise ratio value.

[0117] The second noise estimation result is a noise estimation result under weak smoothing constraints, which can quickly respond to noise changes. The second real-time posterior signal-to-noise ratio is the ratio of the power spectrum of the noisy speech in the current frame to the second noise estimation result (weakly smoothed noise power spectrum), which is consistent with the definition of the first real-time posterior signal-to-noise ratio. However, due to the weak smoothing characteristics of the second noise estimation result, the second real-time posterior signal-to-noise ratio can follow noise abrupt changes and the switching between speech and noise more quickly, and has a better response speed.

[0118] The first and second preset values ​​are fixed parameters obtained in advance based on a large number of noise scenarios (stationary noise, abrupt noise, mixed noise, etc.), and do not require dynamic adjustment. Their function is to provide a benchmark threshold for mapping calculation, ensuring the response speed and accuracy of the judgment of the probability of the second speech presence. For example, the first preset value can be set to 0.5 (i.e., the benchmark for the probability of the speech not existing is 50%), and the second preset value can be set to 15dB (i.e., the benchmark for the prior signal-to-noise ratio is 15dB), which is suitable for most practical application scenarios. Optionally, the mapping formula for the probability of the second speech presence can be expressed as:

[0119]

[0120] This application provides a signal denoising method that extracts the real-time speech absence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio from the first noise estimation result and performs multi-dimensional mapping calculation to obtain the first speech presence probability. It also extracts the second real-time posterior signal-to-noise ratio from the second noise estimation result and combines it with a fixed preset value to quickly map and obtain the second speech presence probability. This differentiated extraction strategy utilizes both the deep characterization of noise smoothing characteristics of the first path and the rapid response advantage of the second path to abrupt signal-to-noise ratio changes, making the extraction of dual-path speech presence probabilities both accurate and real-time. This provides a more reliable state judgment basis for subsequent dynamic adjustment of the noise estimation update speed.

[0121] As an optional implementation, based on any of the above embodiments, determining whether the first noise estimation result is in a near-stagnant state according to the first speech presence probability and the second speech presence probability includes the following steps:

[0122] First, calculate the whole-frame distribution distance between the probability of the first speech and the probability of the second speech.

[0123] Specifically, the probability of the first speech and the probability of the second speech are both probability distributions corresponding to the entire frame signal. Each frame signal contains multiple frequency points, and each frequency point corresponds to a speech existence probability value, not a probability value of a single frequency point.

[0124] The whole-frame distribution distance is used to quantify the overall difference between the probability of the first speech and the probability of the second speech at all frequency points within the whole frame. It can comprehensively reflect the judgment deviation of the two noise estimation branches on the noise state of the current frame and avoid misjudgment caused by local frequency point deviation. In this embodiment, the calculation of the whole-frame distribution distance covers all effective frequency points of noisy speech processing and adapts to the frequency domain distribution characteristics of the speech signal.

[0125] Secondly, feature parameters corresponding to the distribution distance of the entire frame are extracted; among them, the feature parameters include at least one of the mean, variance, sub-band mean difference, and Itakula distance.

[0126] It should be noted that the Itakura distance in this application represents the Itakura-Saito distance.

[0127] Specifically, after obtaining the whole-frame distribution distance between the probability of the first speech and the probability of the second speech, feature extraction is performed on the distribution distance. The extracted feature parameters are all core parameters that can characterize the degree of difference between the two probability distributions. Each parameter reflects the judgment deviation of the dual-channel noise estimation from different dimensions.

[0128] Among them, the mean parameter reflects the overall deviation level of the probability of the two speech signals existing within the whole frame, the variance parameter reflects the dispersion of the deviation at each frequency point within the whole frame, the sub-band mean difference parameter calculates the mean deviation for key sub-bands of the speech signal (such as the 0~500Hz low-frequency sub-band), adapting to the scene characteristics where actual noise mutations mostly occur in the low-frequency band, and the Itakula distance quantifies the overall matching degree of the two probability distributions from the perspective of spectral distortion. In this embodiment, a single feature parameter or a combination of multiple feature parameters can be selected for subsequent judgment according to the needs of the actual noise scene, balancing the accuracy of the judgment and the flexibility of the algorithm.

[0129] Finally, the first noise estimation result is determined to be in a near-stagnant state based on the characteristic parameters.

[0130] Specifically, the extracted feature parameters are input into a preset activation function for fusion calculation to obtain a flag value used to determine the near-stagnant state. In this embodiment, the sigmoid activation function is selected as the fusion calculation function, and the specific calculation formula is as follows:

[0131]

[0132] in, This indicates the probability that the first speech sound exists. Let $\frac{ ...

[0133] Optionally, the probability of existence of the extracted first speech is... The probability of the existence of a second voice To prevent stalling, anti-stalling constraints are implemented to avoid noise estimation updates stalling when the probability of speech presence approaches 1, thus providing effective probability values ​​for accurate judgment of potential stalling states. Specifically, the speech presence probability is first smoothed, and then an upper limit constraint is set on the smoothed probability value. The smoothing formula for the speech presence probability is as follows:

[0134]

[0135] in, The probability of existence for the smoothed speech; This is a smoothing coefficient, with a value range of [0,1]. The probability of the existence of unprocessed raw speech, corresponding to or The formula for the piecewise constraint of the upper probability limit is:

[0136]

[0137] in, There is an upper probability limit for speech, typically set to 0.99, after constraint. This is the probability value of the existence of effective speech used for the final determination of near-stagnant states.

[0138] Specifically, the probability of the first speech occurrence based on the anti-stagnation constraint. The probability of the existence of a second voice Calculate the whole-frame distribution distance between the two to determine whether the first noise estimation result is in a near-stagnant state.

[0139] The signal denoising method provided in this application calculates the whole-frame distribution distance of the probability of the first speech and the probability of the second speech and extracts diversified feature parameters such as mean and variance to comprehensively judge the near-stagnation state of the first noise estimation result. It can more comprehensively capture the dynamic change characteristics in the noise estimation process, effectively distinguish between normal smooth state and abnormal stagnation state, avoid misjudgment caused by local fluctuations, and thus provide a more accurate decision basis for subsequent adaptive adjustment of the noise estimation update speed.

[0140] As an optional implementation, based on any of the above embodiments, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit is adjusted according to the judgment result of the near-stagnation state, including the following steps:

[0141] If the first noise estimation result is determined to be in a near-stagnant state, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit is increased.

[0142] Specifically, when flag=1 is determined by the above feature parameters, that is, the first noise estimation result is in a near-stagnant state, the update speed enhancement mechanism of the first noise estimation branch is triggered. In essence, it is to adjust the smoothness coefficient of the noise estimation to speed up the noise update rate.

[0143] In this embodiment, the noise smoothing coefficient of the first noise estimation branch is directly adjusted to the basic smoothing coefficient. (In this embodiment) Taking 0.8), the corresponding noise estimation formula is corrected as follows:

[0144]

[0145] Furthermore, the minimum noise tracking value is corrected and updated to the minimum estimate when speech is absent. The formula for the minimum noise tracking value is:

[0146]

[0147] in, This is the inverse mapping of the probability estimation of speech absence, used to estimate the probability of speech absence in the next frame. This method directly improves the update speed of the first noise estimation branch, enabling it to quickly jump out of the near-stagnation state, approach the true noise level, and avoid the vicious cycle of noise underestimation.

[0148] If it is determined that the first noise estimation result is not in a state of imminent stagnation, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit remains unchanged.

[0149] Specifically, when flag=0 is determined by the feature parameters, meaning the first noise estimation result is not in a near-stagnant state, the first noise estimation branch maintains the original noise estimation update logic, and the noise smoothing coefficient is still dynamically adjusted based on the speech existence probability. The adjustment formula for the noise smoothing coefficient is as follows:

[0150]

[0151] in, This is the adjusted noise smoothing coefficient.

[0152] Specifically, the noise power spectrum is calculated according to... The calculation is performed. At this time, the first noise estimation branch maintains the update speed of the strong smoothness constraint to ensure the stability of the noise estimation result, thereby continuously exerting its performance advantage of strong music noise suppression and avoiding fluctuations in the noise estimation result due to unnecessary speed adjustment, which would result in music noise residue.

[0153] Specifically, in this embodiment, the update speed is controlled using a differentiated strategy. The speed increase is triggered only when a near-stagnant state is detected, while the original update logic is maintained in other states. This ensures fast tracking capability in noise mutation scenarios and avoids over-control in non-mutation scenarios, achieving a precise balance between music noise suppression and noise tracking speed.

[0154] This application provides a signal denoising method that dynamically adjusts the update speed of a first noise estimation result based on a near-stagnant state judgment. When the first noise estimation result is detected to be in a near-stagnant state, the update is accelerated in a timely manner to quickly respond to noise abrupt changes and avoid a vicious cycle of underestimation. In the non-stagnant state, the original update rhythm is maintained to ensure the music noise suppression effect. Thus, a balance between smooth processing and tracking response is achieved through a differentiated control mechanism, improving the adaptability and robustness of noise estimation.

[0155] Figure 3 This is a flowchart illustrating a signal denoising method according to another embodiment of this application, as shown below. Figure 3 As shown, as an optional implementation, based on any of the above embodiments, the time-domain output signal after noise reduction is determined based on the noisy speech audio domain signal and the noise power spectrum, including the following steps:

[0156] S301. Calculate the posterior signal-to-noise ratio based on the amplitude spectrum and noise power spectrum of the noisy speech domain signal.

[0157] Specifically, the posterior signal-to-noise ratio (SNR) is the ratio of the power spectrum of the noisy speech signal in the current frame to the noise power spectrum after update rate adjustment. It reflects the actual intensity ratio of speech components to noise components in the noisy speech of the current frame, and its calculation formula is as follows:

[0158]

[0159] in, Let k be the frequency point k and l be the frame index l, and l be the posterior signal-to-noise ratio. The power spectrum is the square of the amplitude spectrum of the noisy speech signal. This is the regulated noise power spectrum.

[0160] S302. Determine the prior signal-to-noise ratio based on the posterior signal-to-noise ratio.

[0161] Specifically, the decision-guided method is used to calculate the a priori signal-to-noise ratio (SNR). The a priori SNR reflects the ratio of the clean speech power spectrum to the noise power spectrum of the current frame, and is an important parameter for subsequent Wiener gain calculation. Its calculation formula is as follows:

[0162]

[0163] in, For the estimated prior signal-to-noise ratio, The smoothing coefficient ranges from 0 to 1. In this embodiment... Choose a fixed value of 0.98. This is the Wiener gain coefficient of the previous frame. This is used to ensure the non-negativity of the prior signal-to-noise ratio and avoid invalid values ​​in the calculation results.

[0164] S303. The Wiener gain coefficient is calculated based on the prior signal-to-noise ratio.

[0165] Optionally, the classic Wiener gain calculation method is adopted, and the Wiener gain coefficient is calculated based on the prior signal-to-noise ratio. This coefficient is used to weight and correct the noisy speech signal, thereby suppressing noise components and preserving clean speech components. The calculation formula is as follows:

[0166]

[0167] in, For frequency point k, frame index The corresponding Wiener gain coefficient ranges from 0 to 1. The value of the coefficient is positively correlated with the prior signal-to-noise ratio. That is, the higher the proportion of speech components, the closer the Wiener gain coefficient is to 1, thus preserving the speech signal to the greatest extent. The higher the proportion of noise components, the closer the Wiener gain coefficient is to 0, thus suppressing the noise signal to the greatest extent.

[0168] S304. Based on the Wiener gain coefficient and the noisy speech domain signal, a clean speech domain signal is obtained.

[0169] Specifically, the noisy audio domain signal is multiplied point-by-point by the Wiener gain coefficients of the corresponding frequency points and frame indices, and then weighted and filtered to remove noise components to obtain the estimated clean audio domain signal. The calculation formula is as follows: ,in, For pure speech audio domain signals, For noisy audio signals, the point-by-point multiplication method ensures that the noise suppression effect at each frequency point is adapted to the speech and noise distribution characteristics of that frequency point.

[0170] S305. Perform an inverse Fourier transform on the pure speech audio domain signal to obtain the time-domain speech signal.

[0171] Specifically, for the estimated pure speech audio domain signal Performing an N-point Inverse Discrete Fourier Transform (IDFT) converts the clean speech signal in the frequency domain into a speech signal in the time domain, resulting in a frame index of... A frame-domain time-domain speech signal with time point n The conversion from frequency domain noise reduction to time domain signal is completed. In this embodiment, the value of N is consistent with the number of DFT points when converting the noisy speech time domain signal to the frequency domain. 512 points are selected to ensure the consistency of signal conversion.

[0172] S306. Perform frame synthesis processing on the time-domain speech signal to obtain the noise-reduced time-domain output signal.

[0173] Specifically, the frame-domain and time-domain speech signals obtained by inverse Fourier transform Frame synthesis processing, which involves overlapping and adding frames, eliminates the inter-frame signal abruptness introduced during frame segmentation and windowing, concatenating discrete frame-domain time-domain signals into a continuous time-domain signal, ultimately yielding the denoised time-domain output signal. The overlap ratio of the frame synthesis process is consistent with the overlap ratio of the frame subprocessing in the previous noisy speech preprocessing, ensuring that the synthesized time-domain signal is distortion-free and without breaks. Moreover, the noise-reduced time-domain output signal not only effectively suppresses noise but also has no obvious residual music noise. At the same time, there is no obvious noise tracking delay in noise change scenarios, which improves the subjective listening experience of the call speech in noisy environments.

[0174] This application provides a signal denoising method that calculates the posterior signal-to-noise ratio (SNR) based on the noisy speech audio domain signal and noise power spectrum, derives the prior SNR, and generates Wiener gain coefficients to obtain a clean speech audio domain signal. The clean signal is then processed through inverse Fourier transform and frame synthesis to obtain the time-domain output signal. This constructs a complete closed-loop processing chain from frequency domain noise suppression to time-domain signal reconstruction. The dynamic calculation of the Wiener filter gain coefficients effectively balances noise cancellation and speech distortion, while frame synthesis further optimizes speech continuity. Ultimately, while ensuring denoising depth, this method improves the naturalness and intelligibility of the output speech, achieving high-quality speech enhancement.

[0175] Figure 4 This is a schematic diagram of the structure of a signal noise reduction device provided in an embodiment of this application, as shown below. Figure 4 As shown, the signal noise reduction device provided in this embodiment is located in an electronic device. The signal noise reduction device 40 provided in this embodiment includes: an acquisition module 41, a noise processing module 42, a control module 43, a noise power spectrum determination module 44, and a time-domain output signal determination module 45.

[0176] Specifically, the acquisition module 41 is used to acquire the noisy speech domain signal; the noise processing module 42 is used to input the noisy speech domain signal to the dual-channel noise estimation unit for noise processing to obtain a first noise estimation result and a second noise estimation result; wherein, the first noise estimation result is an estimation result used to suppress music noise, and the second noise estimation result is an estimation result used to improve the ability to track noise mutations; the adjustment module 43 is used to adjust the update speed of the corresponding first noise estimation result in the dual-channel noise estimation unit according to the first noise estimation result and the second noise estimation result; the noise power spectrum determination module 44 is used to determine the noise power spectrum based on the first noise estimation result after the update speed adjustment; and the time-domain output signal determination module 45 is used to determine the denoised time-domain output signal based on the noisy speech domain signal and the noise power spectrum.

[0177] Optionally, when the control module 43 controls the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit based on the first noise estimation result and the second noise estimation result, it is specifically used to: extract the first speech presence probability from the first noise estimation result and extract the second speech presence probability from the second noise estimation result; determine whether the first noise estimation result is in a near-stagnant state based on the first speech presence probability and the second speech presence probability; and control the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit based on the determination result of the near-stagnant state.

[0178] Optionally, the control module 43, when extracting the first speech presence probability from the first noise estimation result and the second speech presence probability from the second noise estimation result, specifically performs the following: extracting the real-time speech absence probability, the real-time prior signal-to-noise ratio, and the first real-time posterior signal-to-noise ratio from the first noise estimation result, and performing mapping calculation based on the real-time speech absence probability, the real-time prior signal-to-noise ratio, and the first real-time posterior signal-to-noise ratio to obtain the first speech presence probability; extracting the second real-time posterior signal-to-noise ratio from the second noise estimation result, and performing mapping calculation based on the second real-time posterior signal-to-noise ratio, the first preset value, and the second preset value to obtain the second speech presence probability; wherein, the first preset value is a fixed speech absence probability value, and the second preset value is a fixed prior signal-to-noise ratio value.

[0179] Optionally, when the control module 43 determines whether the first noise estimation result is in a near-stagnant state based on the first speech presence probability and the second speech presence probability, it is specifically used to: calculate the whole-frame distribution distance between the first speech presence probability and the second speech presence probability; extract the feature parameters corresponding to the whole-frame distribution distance; wherein the feature parameters include at least one of mean, variance, sub-band mean difference, and Itakula distance; and determine whether the first noise estimation result is in a near-stagnant state based on the feature parameters.

[0180] Optionally, when the control module 43 controls the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit based on the judgment result of the near-stagnation state, it is specifically used to: increase the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit when it is determined that the first noise estimation result is in a near-stagnation state; and keep the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit unchanged when it is determined that the first noise estimation result is not in a near-stagnation state.

[0181] Optionally, the time-domain output signal determination module 45, when determining the denoised time-domain output signal based on the noisy speech domain signal and the noise power spectrum, specifically performs the following steps: calculating the posterior signal-to-noise ratio (SNR) based on the amplitude spectrum and noise power spectrum of the noisy speech domain signal; determining the prior SNR based on the posterior SNR; calculating the Wiener gain coefficient based on the prior SNR; obtaining the clean speech domain signal based on the Wiener gain coefficient and the noisy speech domain signal; performing an inverse Fourier transform on the clean speech domain signal to obtain the time-domain speech signal; and performing frame synthesis processing on the time-domain speech signal to obtain the denoised time-domain output signal.

[0182] It should be noted that the signal noise reduction method provided in the above embodiments can be applied to terminal devices / base stations, or chips or chip modules in terminal devices / base stations.

[0183] This application also provides a chip, which includes at least one processor for executing program instructions to perform a signal noise reduction method as described in the above embodiments.

[0184] This application also provides a chip module, which includes at least one processor for executing program instructions to perform a signal noise reduction method as described in the above embodiments.

[0185] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, as shown below. Figure 5 As shown, the electronic device 50 provided in this embodiment includes: a processor 51 and a memory 52 communicatively connected to the processor 51.

[0186] The memory 52 stores computer execution instructions; the processor 51 executes the computer execution instructions stored in the memory 52 to implement a signal noise reduction method provided in any of the above embodiments.

[0187] The program may include program code, which includes computer-executable instructions. Memory 52 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device.

[0188] In this embodiment, the memory 52 and the processor 51 are connected via a bus. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single straight line, but this does not mean that there is only one bus or one type of bus.

[0189] This application also provides a computer-readable storage medium, which stores computer-executable instructions. When executed by a processor, the computer-executable instructions are used to implement a signal noise reduction method provided in any of the above embodiments.

[0190] This application also provides a computer program product, including a computer program that, when executed by a processor, implements a signal noise reduction method provided in any of the above embodiments.

[0191] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0192] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0193] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0194] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0195] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0196] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0197] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0198] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0199] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A signal noise reduction method, characterized in that, include: Acquire noisy speech domain signals; The noisy speech domain signal is input to a dual-channel noise estimation unit for noise processing to obtain a first noise estimation result and a second noise estimation result; wherein, the first noise estimation result is an estimation result used to suppress music noise, and the second noise estimation result is an estimation result used to improve the ability to track noise abrupt changes; Based on the first noise estimation result and the second noise estimation result, the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result is adjusted. The noise power spectrum is determined based on the first noise estimation result after update speed adjustment; Based on the noisy speech domain signal and the noise power spectrum, the denoised time-domain output signal is determined.

2. The method according to claim 1, characterized in that, The step of adjusting the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result based on the first noise estimation result and the second noise estimation result includes: Extract the first speech presence probability from the first noise estimation result, and extract the second speech presence probability from the second noise estimation result; Based on the probability of the first speech occurrence and the probability of the second speech occurrence, determine whether the first noise estimation result is in a near-stagnant state; Based on the judgment result of the near-stagnation state, the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result is adjusted.

3. The method according to claim 2, characterized in that, Extracting the probability of the presence of a first speech from the first noise estimation result and extracting the probability of the presence of a second speech from the second noise estimation result includes: The real-time speech absence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio are extracted from the first noise estimation result. Based on the real-time speech absence probability, real-time prior signal-to-noise ratio, and first real-time posterior signal-to-noise ratio, a mapping calculation is performed to obtain the first speech presence probability. The second real-time posterior signal-to-noise ratio is extracted from the second noise estimation result. Based on the second real-time posterior signal-to-noise ratio, the first preset value and the second preset value are mapped and calculated to obtain the probability of the second speech presence. The first preset value is a fixed probability value of the speech not being present, and the second preset value is a fixed prior signal-to-noise ratio value.

4. The method according to claim 2, characterized in that, The step of determining whether the first noise estimation result is in a near-stagnant state based on the first speech presence probability and the second speech presence probability includes: Calculate the whole-frame distribution distance between the probability of the first speech occurrence and the probability of the second speech occurrence; Extract the feature parameters corresponding to the whole frame distribution distance; wherein, the feature parameters include at least one of mean, variance, sub-band mean difference, and Itakula distance; Based on the characteristic parameters, determine whether the first noise estimation result is in a near-stagnant state.

5. The method according to claim 2, characterized in that, The step of adjusting the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result based on the judgment result of the near-stagnation state includes: If the first noise estimation result is determined to be in a near-stagnant state, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit is increased; If it is determined that the first noise estimation result is not in a near-stagnant state, the update speed of the corresponding first noise estimation result in the dual-path noise estimation unit remains unchanged.

6. The method according to any one of claims 1-5, characterized in that, The step of determining the denoised time-domain output signal based on the noisy speech domain signal and the noise power spectrum includes: The posterior signal-to-noise ratio is calculated based on the amplitude spectrum of the noisy speech domain signal and the noise power spectrum. The prior signal-to-noise ratio is determined based on the posterior signal-to-noise ratio; The Wiener gain coefficient is calculated based on the prior signal-to-noise ratio. Based on the Wiener gain coefficient and the noisy speech domain signal, a clean speech domain signal is obtained; Perform an inverse Fourier transform on the pure speech audio domain signal to obtain a time-domain speech signal; The time-domain speech signal is subjected to frame synthesis processing to obtain a denoised time-domain output signal.

7. A signal noise reduction device, characterized in that, include: The acquisition module is used to acquire noisy speech audio domain signals; The noise processing module is used to input the noisy speech domain signal to the dual-channel noise estimation unit for noise processing to obtain a first noise estimation result and a second noise estimation result; wherein, the first noise estimation result is an estimation result used to suppress music noise, and the second noise estimation result is an estimation result used to improve the ability to track noise abrupt changes; The control module is used to control the update speed of the dual-path noise estimation unit corresponding to the first noise estimation result based on the first noise estimation result and the second noise estimation result. The noise power spectrum determination module is used to determine the noise power spectrum based on the first noise estimation result after the update rate adjustment. The time-domain output signal determination module is used to determine the denoised time-domain output signal based on the noisy speech domain signal and the noise power spectrum.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1-6.

Images