A high-dynamic-environment intercom voice enhancement method and system based on intelligent noise reduction

Abstract

The application provides a high-dynamic-environment intercom voice enhancement method and system based on intelligent noise reduction. In response to a release event of a push-to-talk button, a two-state noise dictionary is constructed based on impulsive noise components and non-stationary noise components in background noise in a current high-dynamic environment. In response to a press event of the push-to-talk button, when it is detected that there is a transient region matching the impulsive noise components in the noisy intercom audio signal, online updating of the two-state noise dictionary is triggered to obtain an online updating dictionary. The noisy intercom audio signal is reconstructed based on the online updating dictionary to obtain an initial enhanced voice signal. Envelope reconstruction is performed on the voice segment with abnormal zero-crossing rate in the initial enhanced voice signal to generate a final enhanced voice signal. The technical scheme provided by the application can enhance conversation voice in a high-dynamic environment with non-stationary noise and impulsive noise.

Description

Technical Field

[0001] This application relates to the field of speech enhancement technology, and more specifically, to a method and system for enhancing intercom speech in high dynamic environments based on intelligent noise reduction. Background Technology

[0002] Speech enhancement aims to extract clean speech signals from noisy speech, improving speech clarity and intelligibility. In real acoustic environments, target speech is inevitably contaminated by environmental noise, reverberation, and interfering speech, seriously affecting human-computer interaction, communication quality, and the performance of hearing aids. In recent years, with the continuous development of speech enhancement, it is possible to more accurately model the complex distribution of speech and noise, effectively preserving the subtle structure of speech while suppressing noise, thus effectively ensuring the quality of voice communication and interaction.

[0003] In existing speech enhancement methods, the first step is to analyze the noisy speech and estimate its noise or speech components. Then, based on the estimation results, noise is suppressed or subtracted from the original signal using specific rules to highlight the speech. Traditional methods are largely based on signal processing and statistical modeling. However, current mainstream deep learning methods shift to a data-driven, end-to-end learning approach. This involves building a deep neural network to learn a time-frequency mask that can distinguish between speech and noise to achieve speech enhancement. However, in high-dynamic environments, especially in cycling environments, background noise includes not only continuous non-stationary noise (such as vehicle noise and wind noise) but also a large amount of sudden, impact noise (such as the sound of bumps when driving over a curb or the honking of nearby cars). Impulsive noise, such as the stuttering sound of a chain, is characterized by its significant transient nature, high energy, and diverse morphologies. Existing speech enhancement techniques typically employ sparse decomposition or spectral subtraction based on a fixed noise dictionary that has been pre-trained or pre-estimated. However, due to the suddenness and unpredictability of impulsive noise, its waveform and frequency domain distribution often differ from the atoms in the "dual-state noise dictionary" collected during communication intervals (i.e., after the "talk" button is pressed and released). During the duration of a call, as the acoustic environment dynamically evolves, the real-time morphology of the impulsive noise rapidly deviates from the representation range of the initial dictionary, resulting in speech distortion in the initially enhanced dialogue. Therefore, how to enhance dialogue in a highly dynamic environment where non-stationary noise and impulsive noise coexist has become a challenge for the industry. Summary of the Invention

[0004] This application provides a method and system for enhancing intercom voice in high dynamic environments based on intelligent noise reduction, which can enhance conversational voice in high dynamic environments where non-stationary noise and impulsive noise exist.

[0005] In a first aspect, this application provides a method for enhancing intercom voice in high dynamic range environments based on intelligent noise reduction, comprising the following steps:

[0006] During the intermittent period of intercom communication, in response to the release event of pressing the talk button, the impulsive noise component is separated from the background noise of the current high dynamic environment and a dual-state noise dictionary is constructed. The dual-state noise dictionary is used to characterize the composite structure characteristics of impulsive noise and non-stationary noise in the current high dynamic environment.

[0007] In response to the press event of the press-to-talk button, when a transient region matching the impulsive noise component is detected in the noisy intercom audio signal of the current target voice, it is determined to be an impulsive noise event, and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call.

[0008] The noisy intercom audio signal is reconstructed based on the online updated dictionary to obtain an initial enhanced speech signal;

[0009] Zero-crossing rate detection is performed on the speech segments in the initial enhanced speech signal corresponding to the moment of the impulsive noise event. Envelope reconstruction is performed on the speech segments with zero-crossing rate anomalies to generate the final enhanced speech signal.

[0010] In some embodiments, separating the impulsive noise component from the background noise of the current high-dynamic environment and constructing a two-state noise dictionary specifically includes:

[0011] Obtain the background noise of the current high dynamic environment, and separate the impulsive noise component and the non-stationary noise component from the background noise of the current high dynamic environment;

[0012] A two-state noise dictionary is constructed based on the impulsive noise component and the non-stationary noise component.

[0013] In some embodiments, constructing a two-state noise dictionary based on the impulsive noise component and the non-stationary noise component specifically includes:

[0014] Based on the impulsive noise component and the non-stationary noise component, impulsive dictionary atoms for characterizing impulsive noise and stationary dictionary atoms for characterizing non-stationary noise are constructed.

[0015] The impact dictionary atoms and the stable dictionary atoms are cascaded and combined to form a two-state noise dictionary with a composite structure.

[0016] In some embodiments, when a transient region matching an impulsive noise component is detected in the noisy intercom audio signal of the current target speech, it is determined to be an impulsive noise event, and an online update of the dual-state noise dictionary is triggered, specifically including:

[0017] Acquire the noisy intercom audio signal of the current target speech, perform saliency detection on the real-time spectrogram, and extract the transient region in the real-time spectrogram where the energy jump rate exceeds a preset threshold;

[0018] The transient region is matched with the impulsive noise components in the dual-state noise dictionary. If the match is successful, it is determined that there is an impulsive noise event and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current speech environment.

[0019] In some embodiments, the online update of the two-state noise dictionary specifically includes:

[0020] Record the arrival time of the impact noise event, locate and extract the short background noise segment that does not contain the target speech from the buffer of the noisy intercom audio signal, the moment before the arrival time.

[0021] The short-term background noise fragment is used to perform similarity replacement on the corresponding shock dictionary atoms in the dual-state noise dictionary to generate an online updated dictionary.

[0022] In some embodiments, reconstructing the noisy intercom audio signal based on the online updated dictionary to obtain the initial enhanced speech signal specifically includes:

[0023] Based on the online updated dictionary, the noisy intercom audio signal is sparsely decomposed to obtain the sparse coefficient vector of the speech components.

[0024] The initial enhanced speech signal is reconstructed using the sparse coefficient vector of the speech components.

[0025] In some embodiments, background noise of the current high dynamic environment is acquired via a microphone array.

[0026] Secondly, this application provides a high dynamic range intercom voice enhancement system based on intelligent noise reduction, used to perform a high dynamic range intercom voice enhancement method based on intelligent noise reduction, the system comprising:

[0027] A construction module is used to separate impulsive noise components from the background noise of the current high-dynamic environment and construct a dual-state noise dictionary in response to the release event of the press-to-talk button during the intermittent period of intercom communication. The dual-state noise dictionary is used to characterize the composite structure features of impulsive noise and non-stationary noise in the current high-dynamic environment.

[0028] The update module is used to respond to the press event of the press-to-talk button. When it detects that there is a transient region in the noisy intercom audio signal of the current target voice that matches the impulsive noise component, it is determined to be an impulsive noise event and triggers the online update of the dual-state noise dictionary to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call.

[0029] The reconstruction module is also used to reconstruct the noisy intercom audio signal based on the online updated dictionary to obtain an initial enhanced speech signal;

[0030] The generation module is used to perform zero-crossing rate detection on the speech segments corresponding to the moment of the impulsive noise event in the initial enhanced speech signal, perform envelope reconstruction on the speech segments with zero-crossing rate anomalies, and generate the final enhanced speech signal.

[0031] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described method for enhancing high dynamic environment intercom voice based on intelligent noise reduction.

[0032] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for enhancing intercom voice in a high dynamic environment based on intelligent noise reduction.

[0033] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0034] The high dynamic environment intercom voice enhancement method and system based on intelligent noise reduction provided in this application firstly, during the intercom communication interval, in response to the release event of pressing the talk button, impulsive noise components are separated from the background noise of the current high dynamic environment, and a dual-state noise dictionary is constructed. The dual-state noise dictionary is used to characterize the composite structure features of impulsive noise and non-stationary noise in the current high dynamic environment. Secondly, in response to the pressing event of the talk button, when a transient region matching the impulsive noise component is detected in the noisy intercom audio signal of the current target voice, it is determined to be an impulsive noise event, and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call. Then, the noisy intercom audio signal is reconstructed based on the online updated dictionary to obtain an initial enhanced voice signal. Finally, zero-crossing rate detection is performed on the voice segments in the initial enhanced voice signal corresponding to the moment of the impulsive noise event, and envelope reconstruction is performed on the voice segments with zero-crossing rate anomalies to generate the final enhanced voice signal.

[0035] Therefore, this application can enhance conversational speech in high-dynamic environments where non-stationary noise and impulsive noise exist. First, by separating impulsive and non-stationary noise components from background noise and constructing a dual-state noise dictionary during the intermittent periods of intercom communication, the composite noise structure of the environment can be pre-captured during the silent periods of the call interval. This decouples and models two types of noise with different physical characteristics, providing a set of prior basis functions that can accurately characterize the noise features of the current high-dynamic environment for subsequent processing. Second, by detecting transient regions in the noisy intercom audio signal that match the atoms in the impulsive dictionary in real time during the call and triggering online dictionary updates, instantaneous perception of sudden impulsive noise events and dynamic adaptation of the dictionary can be achieved. The dictionary is replaced with similarity features using the clean speech features just before the impulsive event, solving the problem of mismatch between the pre-trained dictionary and the real-time acoustic environment, and ensuring that the atoms used in subsequent sparse decomposition are highly correlated with the current noise waveform. Then, by reconstructing the noisy intercom audio signal based on the online updated dictionary, an initial enhanced speech signal is obtained, which is then used to enhance the speech using the updated dictionary. The sparse representation capability of the signal separates speech components from strong background noise, maintaining the integrity of the speech subject while suppressing non-stationary noise. This avoids the problem of speech distortion in the initially enhanced dialogue speech caused by the real-time form of impact noise deviating from the representation range of the initial dictionary during the dynamic evolution of the acoustic environment. This achieves the initial filtering of coupled noise in high dynamic environments. Finally, by performing zero-crossing rate detection and envelope reconstruction on the speech segments corresponding to the moment of the impact noise event in the initial enhanced speech signal, secondary fine repair is performed on the impact traces that may remain after sparse decomposition. The zero-crossing rate anomaly is used to accurately locate the contaminated segments and restore their natural waveform through envelope reconstruction, eliminating the auditory discomfort caused by transient pulse residue. Finally, a pure and natural enhanced speech signal is output, which significantly improves the intelligibility of intercom speech in noisy environments while maintaining the naturalness of the speech. This enhances the speech of the target dialogue speech under the influence of non-stationary noise and impact noise. In summary, the technical solution provided in this application can enhance dialogue speech in high dynamic environments where non-stationary noise and impact noise exist. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of an application scenario architecture for a high dynamic environment intercom voice enhancement method based on intelligent noise reduction, according to some embodiments of this application.

[0037] Figure 2 This is an exemplary flowchart of a high dynamic range intercom voice enhancement method based on intelligent noise reduction, according to some embodiments of this application.

[0038] Figure 3This is an exemplary flowchart illustrating the determination of an online dictionary update according to some embodiments of this application;

[0039] Figure 4 This is a schematic diagram of the structure of a high dynamic environment intercom voice enhancement system based on intelligent noise reduction, according to some embodiments of this application;

[0040] Figure 5 This is a schematic diagram of the structure of a computer device that implements a high dynamic environment intercom voice enhancement method based on intelligent noise reduction, according to some embodiments of this application. Detailed Implementation

[0041] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] refer to Figure 1 This figure is a schematic diagram of an application scenario architecture of a high dynamic environment intercom voice enhancement method based on intelligent noise reduction according to some embodiments of this application. The left side of the figure shows the process of collecting environmental background noise and separating impulsive noise and non-stationary noise to construct a dual-state noise dictionary during the communication interval when the helmet intercom is released after pressing the talk button in a high dynamic cycling scenario. The middle part of the figure shows the process of the helmet intercom collecting noisy intercom audio after the talk button is pressed, detecting impulsive noise events in real time and updating the noise dictionary online. The right side of the figure shows the process of finally outputting a clear enhanced voice signal through sparse decomposition reconstruction, zero-crossing rate detection and envelope reconstruction processing.

[0043] refer to Figure 2 This figure is an exemplary flowchart of a high dynamic environment intercom voice enhancement method based on intelligent noise reduction according to some embodiments of this application. The figure mainly includes the following steps:

[0044] In step S101, during the intercom communication interval, in response to the release event of pressing the talk button, the impulsive noise component is separated from the background noise of the current high dynamic environment and a dual-state noise dictionary is constructed. The dual-state noise dictionary is used to characterize the composite structure features of impulsive noise and non-stationary noise in the current high dynamic environment.

[0045] It should be noted that the "press-to-talk" button in this application refers to a button switch on a walkie-talkie device used to control the switching between voice transmission and reception states, such as the "press-to-talk" button on a motorcycle helmet walkie-talkie. This button operates on a press-to-conduct and release-to-disconnect principle. When the user presses the button, the device enters the voice transmission state, and the voice signal collected by the microphone can be transmitted to the other end through the communication link. When the user releases the button, the device exits the voice transmission state and enters the reception state or mute state. At this time, the microphone is only used to collect ambient noise and does not transmit voice, thereby realizing voice communication control in half-duplex mode. This button provides a clear trigger event basis for distinguishing between the voice call period and the ambient noise collection period.

[0046] In some embodiments, separating the impulsive noise component from the background noise of the current high-dynamic environment and constructing a two-state noise dictionary is achieved through the following steps:

[0047] Obtain the background noise of the current high dynamic environment, and separate the impulsive noise component and the non-stationary noise component from the background noise of the current high dynamic environment;

[0048] A two-state noise dictionary is constructed based on the impulsive noise component and the non-stationary noise component.

[0049] In some embodiments, the acquisition of the background noise of the current high-dynamic environment and the separation of impulsive noise components and non-stationary noise components from the background noise of the current high-dynamic environment are achieved by the following steps:

[0050] Obtain the background noise of the current high dynamic environment;

[0051] Perform time-frequency transformation on the background noise to obtain its transient energy fluctuation trajectory in different frequency bands;

[0052] The impulsive noise component and the non-stationary noise component in the background noise are identified based on the transient energy fluctuation trajectory in different frequency bands.

[0053] In specific implementation, firstly, during the intercom communication interval, in response to the release event of pressing the "talk" button, the background noise of the current high dynamic environment is acquired through the microphone array. Specifically, the microphone array connected to the audio codec of the intercom terminal is used to collect the background noise of the current high dynamic environment. The background noise is stored in a buffer in pulse code modulation format as a noise sample to be analyzed. The noise sample to be analyzed is a discrete sequence of the original environmental sound wave signal without any voice activity, after digital quantization. Then, the background noise is subjected to time-frequency transformation to obtain its instantaneous frequency response in different frequency bands. Specifically, the noise sample to be analyzed is segmented into frames to obtain a windowed speech frame sequence. A Hamming window is applied to each windowed speech frame, and its spectrum is obtained through a short-time Fourier transform. The power spectrum matrix is ​​obtained by taking the modulus and square of the spectra of all windowed speech frames. The power spectrum matrices are stacked along the time axis to form a spectrogram, where the horizontal axis represents time, the vertical axis represents frequency, and the pixel values ​​represent energy density. This constitutes the transient energy fluctuation trajectory of the background noise in different frequency bands. Subsequently, the background noise is identified based on the transient energy fluctuation trajectory in different frequency bands. The analysis distinguishes between impulsive and non-stationary noise components in the sound. For the identification of impulsive noise components, the spectrogram is analyzed frame-by-frame. The total energy of each frame across all frequency bands is calculated to obtain a frame energy sequence. First-order difference calculations are performed on the frame energy sequence to obtain an energy change rate sequence. Frames exceeding a preset threshold in the energy change rate sequence are marked as impulsive candidate frames. Connected component merging is performed on these impulsive candidate frames. The time intervals corresponding to consecutively occurring impulsive candidate frames are determined as impulsive event intervals. The time-domain waveforms within the impulsive event intervals are extracted from the noise samples to be analyzed. As an impulsive noise component, the impulsive noise component is a transient pulse signal with a duration shorter than a preset threshold. For the identification of non-stationary noise components, the signal within the impulsive event interval is removed from the noise sample to be analyzed, the remaining noise signal is divided into frequency bands, the variance of energy in each frequency band is calculated, the frequency band with energy variance exceeding a preset fluctuation threshold is marked as a non-stationary active frequency band, the time domain signal in the non-stationary active frequency band is de-wound and its instantaneous frequency change curve is extracted, and the frequency band signal corresponding to the instantaneous frequency change curve is determined as a non-stationary noise component.

[0054] It should be noted that the non-stationary noise component in this application is a non-pulse background interference whose statistical characteristics change slowly over time but are continuously distributed in the frequency domain. In a high-dynamic intercom environment, impulse noise and non-stationary noise exhibit distinctly different physical characteristics in the time and frequency domains. The former manifests as transient pulses with extremely short durations and rapid energy jumps, while the latter manifests as background fluctuations whose statistical characteristics change slowly over time but are continuously distributed in the frequency domain. If the two are not distinguished and a unified processing strategy is adopted, the noise reduction algorithm will be unable to simultaneously achieve accurate removal of impulse noise and effective tracking of non-stationary noise. By separating the two, a rapid response mechanism based on transient detection and waveform restoration can be constructed for impulse noise, and a continuous suppression framework based on dynamic dictionary learning and sparse representation can be constructed for non-stationary noise. This provides a structured prior basis for adaptive noise reduction processing triggered by noise type during the call.

[0055] In some embodiments, the construction of a two-state noise dictionary based on the impulsive noise component and the non-stationary noise component is achieved by the following steps:

[0056] Based on the impulsive noise component and the non-stationary noise component, impulsive dictionary atoms for characterizing impulsive noise and stationary dictionary atoms for characterizing non-stationary noise are constructed.

[0057] The impact dictionary atoms and the stable dictionary atoms are cascaded and combined to form a two-state noise dictionary with a composite structure.

[0058] In specific implementation, firstly, the impulsive noise components are subjected to equal-length frame segmentation to obtain an impulsive noise frame set composed of frames of fixed length. Then, the non-stationary noise components are subjected to equal-length frame segmentation to obtain a stationary noise frame set. Each frame in the impulsive noise frame set is arranged as a column vector to form an impulsive training matrix. The impulsive training matrix is ​​a two-dimensional numerical array with impulsive noise frames as column vectors. K-singular value decomposition dictionary learning is performed on the impulsive training matrix. Specifically, K column vectors in the impulsive training matrix are randomly selected as initial impulsive dictionary atoms to form an impulsive dictionary matrix. Then, in the sparse coding stage, the orthogonal matching pursuit algorithm is used to calculate the impulsive noise. The training matrix contains a sparse representation coefficient matrix for each sample under the impulse dictionary matrix. This sparse representation coefficient matrix is ​​a weighted coefficient array describing how each sample is linearly combined from dictionary atoms. During the dictionary update phase, singular value decomposition is performed on each impulse dictionary atom in the impulse dictionary matrix. This sparse coding and dictionary update process is iteratively executed until convergence, resulting in an impulse dictionary matrix composed of multiple impulse dictionary atoms. Each column of this impulse dictionary matrix represents an impulse dictionary atom used to characterize the impulse noise. Each impulse dictionary atom is a basis function vector capable of sparsely representing the waveform structure of the impulse noise. Simultaneously, each frame in the set of stationary noise frames is used as a column. The vectors are arranged to form a stationary training matrix. K-singular value decomposition (KSVD) dictionary learning is then performed on this stationary training matrix. Specifically, K column vectors from the stationary training matrix are randomly selected as initial stationary dictionary atoms to form a stationary dictionary matrix. In the sparse coding stage, the orthogonal matching pursuit algorithm is used to calculate the sparse representation coefficient matrix of each sample in the stationary training matrix under the stationary dictionary matrix. In the dictionary update stage, singular value decomposition is performed on each atom in the stationary dictionary matrix to update it. This sparse coding and dictionary update process is iteratively executed until convergence, resulting in a stationary dictionary matrix composed of multiple stationary dictionary atoms. Each column in the stationary dictionary matrix is ​​used to represent non-stationary vectors. The stationary dictionary atoms for stationary noise are basis function vectors that can sparsely represent the structure of non-stationary noise waveforms. Finally, the impulse dictionary atoms and the stationary dictionary atoms are cascaded and combined to form a bistate noise dictionary with a composite structure. That is, an impulse dictionary matrix containing the impulse dictionary atoms and a stationary dictionary matrix containing the stationary dictionary atoms are obtained. The impulse dictionary matrix and the stationary dictionary matrix are concatenated in the column direction, that is, the number of columns of the two matrices remains unchanged. All columns of the stationary dictionary matrix are arranged sequentially after all columns of the impulse dictionary matrix to form a new composite dictionary matrix, which is the bistate noise dictionary.

[0059] It should be noted that the dual-state noise dictionary in this application is an overcomplete basis function matrix that simultaneously contains impulse dictionary atoms and stationary dictionary atoms. It is used to characterize the composite structural features of impulse noise and non-stationary noise in the current high-dynamic environment. The impulse dictionary atom set and the stationary dictionary atom set are independently trained based on the separated impulse noise components and non-stationary noise components, respectively. The two are then cascaded and combined to form the dual-state noise dictionary. This ensures that the impulse dictionary atoms focus on capturing the waveform structure of transient pulses, while the stationary dictionary atoms focus on describing the spectral evolution of background fluctuations. Thus, in the subsequent call processing, when an impulse noise event is detected, the impulse dictionary atoms can be directly called for fast atom replacement and waveform repair. In non-impact periods, the stationary dictionary atoms are called for continuous sparse decomposition and background tracking. Ultimately, this achieves the divide-and-conquer processing and precise suppression of the two types of coupled noise in the high-dynamic environment.

[0060] In step S102, in response to the press event of the press-to-talk button, when a transient region matching the impulsive noise component is detected in the noisy intercom audio signal of the current target voice, it is determined to be an impulsive noise event, and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call.

[0061] In some embodiments, reference Figure 3 As shown in the figure, this is an exemplary flowchart of determining an online updated dictionary according to some embodiments of this application. In this embodiment, when a transient region matching an impulsive noise component is detected in the noisy intercom audio signal of the current target speech, it is determined to be an impulsive noise event, and the online update of the dual-state noise dictionary is triggered. This can be achieved by the following steps:

[0062] In step S1021, the noisy intercom audio signal of the current target speech is acquired, and continuous time-frequency analysis is performed on the noisy intercom audio signal to obtain the real-time spectrogram corresponding to the noisy intercom audio signal.

[0063] In step S1022, saliency detection is performed on the real-time spectrogram to extract transient regions in the real-time spectrogram where the energy jump rate exceeds a preset threshold.

[0064] In step S1023, the transient region is matched with the impulsive noise components in the dual-state noise dictionary. If the match is successful, it is determined that there is an impulsive noise event and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current speech environment.

[0065] In specific implementation, firstly, the key scanning circuit of the intercom terminal detects the press event of the "Press and Talk" button and generates a trigger signal. This trigger signal wakes up the audio codec in standby mode and configures its sampling parameters to acquire a noisy intercom audio signal containing the target speech through the microphone array. The noisy intercom audio signal refers to the raw digital audio stream acquired by the microphone array during the intercom communication session. Its physical composition includes the speech signal emitted by the target speaker and high-dynamic environmental background noise superimposed on the speech signal in both the time and frequency domains. Continuous time-frequency analysis is performed on the noisy intercom audio signal to obtain the corresponding real-time signal. The real-time spectrogram is specifically generated by performing frame-by-frame windowing processing on the noisy intercom audio signal to obtain a windowed speech frame sequence. A Hamming window is applied to each windowed speech frame, and a short-time Fourier transform is used to obtain the spectrum of that windowed speech frame. The spectra of each frame are squared and stacked along the time axis to form a real-time spectrogram, which is a matrix describing the distribution of signal energy density in a two-dimensional plane of time and frequency. Then, saliency detection is performed on the real-time spectrogram to extract transient regions where the energy jump rate exceeds a preset threshold. Specifically, the real-time spectrogram is analyzed frame by frame, and the sum of energy in all frequency bands for each frame is calculated to obtain a frame energy sequence. The frame energy sequence is calculated using first-order difference to obtain an energy change rate sequence. The time interval corresponding to frames in the energy change rate sequence whose energy jump rate continuously exceeds a preset threshold is determined as the transient region. Finally, the transient region is matched with the impulsive noise components in the dual-state noise dictionary. If the match is successful, it is determined that there is an impulsive noise event, and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. Specifically, the time-domain waveform segment corresponding to the transient region is extracted from the noisy intercom audio signal as the signal segment to be matched, and the signal segment to be matched is compared with the impulsive noise components in the dual-state noise dictionary. Each impulse dictionary atom corresponding to the noise component calculates its inner product. The maximum value of the inner product is compared with a preset matching threshold. If the maximum value of the inner product exceeds the preset matching threshold, the match is considered successful. This indicates that an impulse noise event exists at the current moment, and the arrival time of the impulse noise event is recorded. The impulse noise event is the moment when a transient pulse highly similar to the impulse dictionary atom appears in the noisy intercom audio signal. After determining the impulse noise event, an online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. The preset matching threshold can be set according to actual needs or expert knowledge, and is not limited here.

[0066] In some embodiments, the online update of the two-state noise dictionary includes:

[0067] Record the arrival time of the impact noise event, locate and extract the short background noise segment that does not contain the target speech from the buffer of the noisy intercom audio signal, the moment before the arrival time.

[0068] The short-term background noise fragment is used to perform similarity replacement on the corresponding shock dictionary atoms in the dual-state noise dictionary to generate an online updated dictionary.

[0069] In specific implementation, firstly, the arrival time of the impact noise event is recorded. Then, a short signal segment preceding the arrival time is located and extracted from the buffer of the noisy intercom audio signal. Simultaneously, a pre-embedded voice activity detection module is used to detect this short signal segment. If it is determined that the short signal segment does not contain the target speech (i.e., it is in the noise acquisition period after the button is released), the short signal segment is identified as a short background noise segment. If it is determined to contain the target speech, a backward search is performed until a noise segment without target speech is found. Then, the short background noise segment is used to analyze the dual-state noise word. The corresponding impact dictionary atoms in the dictionary are replaced based on similarity. Specifically, the inner product of the short-time background noise segment and each impact dictionary atom in the dual-state noise dictionary is calculated. The impact dictionary atom with the largest inner product value is determined as the atom to be replaced that is most similar to the current noise feature. The time-domain waveform vector of the short-time background noise segment is directly used to cover the column vector position of the atom to be replaced. The replaced impact dictionary atom is then normalized using the L2 norm to maintain the energy consistency of the dictionary atoms. The impact dictionary atoms after similarity replacement are then recombined with the original stationary dictionary atoms to generate an online updated dictionary.

[0070] It should be noted that the online updated dictionary in this application refers to the updated dual-state noise dictionary. Specifically, it is a composite dictionary matrix in which the impulse dictionary atoms have been replaced with similarity-based replacements based on the clean speech features before the occurrence of the current impulse noise event, while the stationary dictionary atoms remain unchanged. The online updated dictionary is used to dynamically adapt to the real-time changes in impulse noise during the current call. In a high-dynamic intercom environment, the occurrence of impulse noise is sudden and unpredictable, and its waveform characteristics will deviate from the fixed form of the impulse dictionary atoms in the pre-trained dual-state noise dictionary as the acoustic environment changes in real time (such as speaker position movement or noise source type change). If a static pre-trained dictionary is used directly to sparsely represent the current impulse noise, the matching degree between atoms and noise waveforms will decrease, making it impossible to accurately separate noise components during sparse decomposition. This can lead to residual transient pulse traces or the introduction of new distortions during speech reconstruction. By using an online update mechanism to replace impulse dictionary atoms with similarity segments of short-term background noise before the impulse noise event, the dictionary can dynamically adapt to the transient acoustic features at the current moment. This ensures that the atoms used for subsequent impulse noise suppression have a higher correlation with the real noise waveform, thereby achieving accurate capture and repair of sudden impulses while maintaining the integrity of the main speech content.

[0071] In step S103, the noisy intercom audio signal is reconstructed based on the online updated dictionary to obtain an initial enhanced speech signal.

[0072] In some embodiments, the reconstruction of the noisy intercom audio signal based on the online updated dictionary to obtain the initial enhanced speech signal is achieved through the following steps:

[0073] Based on the online updated dictionary, the noisy intercom audio signal is sparsely decomposed to obtain the sparse coefficient vector of the speech components.

[0074] The initial enhanced speech signal is reconstructed using the sparse coefficient vector of the speech components.

[0075] In specific implementation, firstly, the noisy intercom audio signal is processed by frame segmentation and windowing to obtain a noisy speech frame sequence composed of fixed-length frames. The noisy speech frame sequence is a discrete frame set formed by segmenting and windowing a continuous time-domain signal. Based on the online updated dictionary, the noisy intercom audio signal is sparsely decomposed to obtain the sparse coefficient vector of the speech components. Specifically, an orthogonal matching pursuit algorithm is executed on each frame signal in the noisy speech frame sequence. The current residual signal is initialized as the current frame signal, and the selected atom index set is initialized as an empty set. The iteration begins, and the inner product of the current residual signal and each atom in the online updated dictionary is calculated. The atom with the largest absolute value of the inner product is selected as the matching atom, and its index in the dictionary is added to the selected atom index set. The least squares method is used to solve for the sparse representation coefficients of all atoms corresponding to the current selected atom index set for the current frame signal. These coefficients are the sparse coefficient vector of the speech components of the current frame. The current frame signal is then subtracted from the sum of the values ​​of the currently selected atoms and their corresponding sparse coefficients. The product of the number vectors yields the updated residual signal. This iterative process is repeated until a preset sparsity upper limit is reached or the L2 norm of the residual signal falls below a preset threshold. The output is then the sparse coefficient vector of the speech component in the current frame. This sparse coefficient vector describes how the current frame signal is linearly combined from a small number of atoms in the online update dictionary. Then, the sparse coefficient vectors of the speech components from all frames are combined in frame order into a sparse coefficient vector matrix. This matrix is ​​an array of sparse coefficient vectors corresponding to each frame signal. The initial enhanced speech signal is reconstructed using these sparse coefficient vectors. Specifically, the sparse coefficient vector matrix is ​​multiplied by the online update dictionary to obtain the reconstructed time-domain waveform vector for each frame signal. All reconstructed time-domain waveform vectors are overlapped and added together. That is, the corresponding sampling point values ​​are added for overlapping regions of adjacent frames, while values ​​are directly taken for non-overlapping regions. This restores the signal to a continuous time-domain waveform with the same length as the input signal, generating the initial enhanced speech signal.

[0076] It should be noted that the initial enhanced speech signal in this application is a speech waveform that has been reconstructed by sparse decomposition and the background noise has been suppressed.

[0077] In step S104, zero-crossing rate detection is performed on the speech segments in the initial enhanced speech signal corresponding to the moment of the impulsive noise event, and envelope reconstruction is performed on the speech segments with zero-crossing rate anomalies to generate the final enhanced speech signal.

[0078] In some embodiments, the following steps are used to perform zero-crossing rate detection on the speech segments corresponding to the moment of the impulsive noise event in the initial enhanced speech signal, and to perform envelope reconstruction on the speech segments with zero-crossing rate anomalies to generate the final enhanced speech signal:

[0079] Extract the arrival time of the impact noise event, and locate the impact-contaminated speech segment in the initial enhanced speech signal based on the arrival time;

[0080] Calculate the actual zero-crossing rate of the impact-contaminated speech segment and compare the actual zero-crossing rate with the historical zero-crossing rate of the previous speech frame of the impact-contaminated speech segment. If the actual zero-crossing rate exceeds the historical zero-crossing rate by a preset multiple, it is determined that the segment has an abnormal zero-crossing rate.

[0081] Extract the amplitude envelope of the previous speech frame of the impact-contaminated speech segment as a reference envelope, and multiply the reference envelope by a preset energy attenuation factor to generate the target reconstruction envelope;

[0082] The target reconstruction envelope is used to replace the waveform envelope of the impact-contaminated speech segment to generate a speech segment with impact trace repair.

[0083] The audio segments repaired by the impact marks are spliced ​​back to their corresponding positions in the initial enhanced audio signal to generate the final enhanced audio signal.

[0084] In specific implementation, firstly, the arrival time of the impact noise event is extracted. Based on the arrival time, the impact-contaminated speech segment is located on the time-domain waveform of the initial enhanced speech signal. Specifically, a fixed-length time window is extended forward and backward from the arrival time, and the sampling point sequence within this time window is extracted from the initial enhanced speech signal as the impact-contaminated speech segment. The impact-contaminated speech segment is a local time-domain waveform in the initial enhanced speech signal that may contain residual transient pulse traces. Then, the actual zero-crossing rate of the impact-contaminated speech segment is calculated. Specifically, the sign of two adjacent sampling points in the impact-contaminated speech segment is determined; if the signs are opposite, then... Each zero-crossing is counted as one occurrence. The actual zero-crossing rate is obtained by counting the number of zero-crossings within a unit length. This actual zero-crossing rate is then compared with the historical zero-crossing rate of the preceding speech frame of the impact-contaminated speech segment. The historical zero-crossing rate of the preceding speech frame is the zero-crossing rate of the complete speech frame immediately preceding the impact-contaminated speech segment, calculated using the same method. If the actual zero-crossing rate exceeds a preset multiple (e.g., twice, not limited here) of the historical zero-crossing rate, the segment is determined to have an abnormal zero-crossing rate. This abnormal zero-crossing rate indicates that the segment has experienced high-frequency oscillations or waveform distortion due to residual impact noise. Subsequently, the amplitude envelope of the preceding speech frame of the impact-contaminated speech segment is extracted as a reference envelope. Specifically, for... The time-domain waveform of the previous speech frame is subjected to Hilbert transform to obtain an analytic signal. The modulus of the analytic signal is then taken to obtain an instantaneous amplitude sequence. This instantaneous amplitude sequence is smoothed and filtered to remove glitches, resulting in a reference envelope. This reference envelope is a contour curve describing the amplitude change of the previous speech frame over time. The reference envelope is multiplied by a preset energy attenuation factor to generate a target reconstructed envelope. The energy attenuation factor is a constant less than 1 used to simulate the natural energy decay trend after an impact. The target reconstructed envelope is used to reconstruct the target contour curve of the amplitude change of the impact-contaminated segment. Next, the target reconstructed envelope is used to perform waveform envelope replacement on the impact-contaminated speech segment. Specifically, the impact-contaminated speech segment... The audio segment undergoes Hilbert transform to obtain its analytic signal. The instantaneous phase is extracted from the analytic signal, and the instantaneous amplitude of the analytic signal is replaced with the amplitude of the target reconstructed envelope at the corresponding moment. The real part of the replaced analytic signal is then taken to obtain the impact-trace-repaired audio segment. The impact-trace-repaired audio segment is a time-domain waveform in which the instantaneous phase remains unchanged and the envelope has been reconstructed. Finally, the impact-trace-repaired audio segment is spliced ​​back to the corresponding position in the initial enhanced audio signal. Specifically, the sampling point sequence of the impact-trace-repaired audio segment is used to cover the original sampling points in the time interval of the impact-contaminated audio segment in the initial enhanced audio signal to generate the final enhanced audio signal.

[0085] It should be noted that the final enhanced speech signal in this application refers to the speech signal composed of all sampling points arranged in chronological order after the impact traces have been repaired.

[0086] Furthermore, in another aspect of this application, in some embodiments, this application provides a high dynamic range intercom voice enhancement system based on intelligent noise reduction, referencing... Figure 4 The figure is a schematic diagram of the structure of a high dynamic environment intercom voice enhancement system based on intelligent noise reduction according to some embodiments of this application. The high dynamic environment intercom voice enhancement system based on intelligent noise reduction includes: a construction module 201, an update module 202, a reconstruction module 203, and a generation module 204, which are described below:

[0087] Construction module 201, in this application, is mainly used to separate the impulsive noise component from the background noise of the current high dynamic environment and construct a dual-state noise dictionary in response to the release event of pressing the talk button during the intermittent period of intercom communication. The dual-state noise dictionary is used to characterize the composite structure characteristics of impulsive noise and non-stationary noise in the current high dynamic environment.

[0088] The update module 202 in this application is mainly used to respond to the press event of the press-to-talk button. When a transient region matching the impulsive noise component is detected in the noisy intercom audio signal of the current target voice, it is determined to be an impulsive noise event, and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call.

[0089] The reconstruction module 203 is also used to reconstruct the noisy intercom audio signal based on the online updated dictionary to obtain an initial enhanced speech signal;

[0090] The generation module 204 is mainly used to perform zero-crossing rate detection on the speech segments corresponding to the moment of the impulsive noise event in the initial enhanced speech signal, perform envelope reconstruction on the speech segments with zero-crossing rate anomalies, and generate the final enhanced speech signal.

[0091] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described method for enhancing high dynamic environment intercom voice based on intelligent noise reduction.

[0092] In some embodiments, reference Figure 5The figure is a schematic diagram of the structure of a computer device implementing a high dynamic range intercom voice enhancement method based on intelligent noise reduction, according to some embodiments of this application. The high dynamic range intercom voice enhancement method based on intelligent noise reduction in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0093] The processor 301 can be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the intelligent noise reduction-based high dynamic environment intercom voice enhancement method in this application.

[0094] The communication bus 302 can be used to transmit information between the aforementioned components.

[0095] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0096] The memory 303 stores program code for executing the solution of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the high dynamic environment intercom voice enhancement method based on intelligent noise reduction can be achieved by the processor 301 and one or more software modules in the program code in the memory 303.

[0097] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0098] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0099] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0100] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for enhancing intercom voice in high dynamic environments based on intelligent noise reduction.

[0101] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0102] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for enhancing intercom voice in high dynamic range environments based on intelligent noise reduction, characterized in that, Includes the following steps: During the intermittent period of intercom communication, in response to the release event of pressing the talk button, the impulsive noise component is separated from the background noise of the current high dynamic environment and a dual-state noise dictionary is constructed. The dual-state noise dictionary is used to characterize the composite structure characteristics of impulsive noise and non-stationary noise in the current high dynamic environment. In response to the press event of the press-to-talk button, when a transient region matching the impulsive noise component is detected in the noisy intercom audio signal of the current target voice, it is determined to be an impulsive noise event, and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call. The noisy intercom audio signal is reconstructed based on the online updated dictionary to obtain an initial enhanced speech signal; Zero-crossing rate detection is performed on the speech segments in the initial enhanced speech signal corresponding to the moment of the impulsive noise event. Envelope reconstruction is performed on the speech segments with zero-crossing rate anomalies to generate the final enhanced speech signal.

2. The method as described in claim 1, characterized in that, The process of separating the impact noise component from the background noise of the current high-dynamic environment and constructing a two-state noise dictionary specifically includes: Obtain the background noise of the current high dynamic environment, and separate the impulsive noise component and the non-stationary noise component from the background noise of the current high dynamic environment; A two-state noise dictionary is constructed based on the impulsive noise component and the non-stationary noise component.

3. The method as described in claim 2, characterized in that, The two-state noise dictionary constructed based on the impulsive noise component and the non-stationary noise component specifically includes: Based on the impulsive noise component and the non-stationary noise component, impulsive dictionary atoms for characterizing impulsive noise and stationary dictionary atoms for characterizing non-stationary noise are constructed. The impact dictionary atoms and the stable dictionary atoms are cascaded and combined to form a two-state noise dictionary with a composite structure.

4. The method as described in claim 1, characterized in that, When a transient region matching an impulsive noise component is detected in the noisy intercom audio signal of the current target speech, it is determined to be an impulsive noise event, and an online update of the dual-state noise dictionary is triggered, specifically including: Acquire the noisy intercom audio signal of the current target speech, perform continuous time-frequency analysis on the noisy intercom audio signal, and obtain the real-time spectrogram corresponding to the noisy intercom audio signal; Saliency detection is performed on the real-time spectrogram to extract transient regions in the real-time spectrogram where the energy jump rate exceeds a preset threshold; The transient region is matched with the impulsive noise components in the dual-state noise dictionary. If the match is successful, it is determined that there is an impulsive noise event and the online update of the dual-state noise dictionary is triggered to obtain an online updated dictionary adapted to the current speech environment.

5. The method as described in claim 4, characterized in that, The online update of the two-state noise dictionary specifically includes: Record the arrival time of the impact noise event, locate and extract the short background noise segment that does not contain the target speech from the buffer of the noisy intercom audio signal, the moment before the arrival time. The short-term background noise fragment is used to perform similarity replacement on the corresponding shock dictionary atoms in the dual-state noise dictionary to generate an online updated dictionary.

6. The method as described in claim 1, characterized in that, The initial enhanced speech signal is reconstructed based on the online updated dictionary to obtain the noisy intercom audio signal, specifically including: Based on the online updated dictionary, the noisy intercom audio signal is sparsely decomposed to obtain the sparse coefficient vector of the speech components. The initial enhanced speech signal is reconstructed using the sparse coefficient vector of the speech components.

7. The method as described in claim 2, characterized in that, Background noise in the current high-dynamic environment is collected using a microphone array.

8. A high dynamic range intercom voice enhancement system based on intelligent noise reduction, used to execute the high dynamic range intercom voice enhancement method based on intelligent noise reduction as described in any one of claims 1 to 7, characterized in that, The system includes: A construction module is used to separate impulsive noise components from the background noise of the current high-dynamic environment and construct a dual-state noise dictionary in response to the release event of the press-to-talk button during the intermittent period of intercom communication. The dual-state noise dictionary is used to characterize the composite structure features of impulsive noise and non-stationary noise in the current high-dynamic environment. The update module is used to respond to the press event of the press-to-talk button. When it detects that there is a transient region in the noisy intercom audio signal of the current target voice that matches the impulsive noise component, it is determined to be an impulsive noise event and triggers the online update of the dual-state noise dictionary to obtain an online updated dictionary adapted to the current voice environment. The online updated dictionary is used to dynamically adapt to the real-time changes of impulsive noise during the current call. The reconstruction module is also used to reconstruct the noisy intercom audio signal based on the online updated dictionary to obtain an initial enhanced speech signal; The generation module is used to perform zero-crossing rate detection on the speech segments corresponding to the moment of the impulsive noise event in the initial enhanced speech signal, perform envelope reconstruction on the speech segments with zero-crossing rate anomalies, and generate the final enhanced speech signal.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the high dynamic range intercom voice enhancement method based on intelligent noise reduction as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the high dynamic environment intercom voice enhancement method based on intelligent noise reduction as described in any one of claims 1 to 7.

Images