A narrowband satellite voice communication noise reduction system and method
By preprocessing, separating, and performing AI completion on narrowband satellite voice signals, the problem of reduced signal-to-noise ratio in narrowband satellite communication was solved, thus improving communication quality and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SEVEN STAR COMM TECH (BEIJING) CO LTD
- Filing Date
- 2025-12-17
- Publication Date
- 2026-05-12
Smart Images
Figure CN121617408B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of satellite communication technology, and more specifically, to a narrowband satellite voice communication noise reduction system and method. Background Technology
[0002] In satellite voice communication scenarios, the core communication channel has a critical characteristic of bandwidth limitation (1.2K, 2.4K, 4.8K), which inherently restricts the transmission bandwidth of voice signals. At the same time, satellite communication links are susceptible to factors such as space electromagnetic interference, signal attenuation, and multipath effects, resulting in a significant reduction in the signal-to-noise ratio (SNR) during communication. Summary of the Invention
[0003] This invention provides a narrowband satellite voice communication noise reduction system and method to solve the problem of poor communication quality in existing satellite voice communication under narrowband transmission constraints, including:
[0004] The preprocessing module is used to acquire the raw speech signal from the satellite receiver, preprocess the raw speech signal, and obtain the preprocessed speech signal.
[0005] The separation module is used to establish a speech separation model and separate the effective speech signal and noise signal from the preprocessed speech signal according to the speech separation model;
[0006] The enhancement module is used to build an AI completion model, and fill in the abnormalities of the effective speech signal according to the AI completion model to obtain the enhanced speech signal.
[0007] The post-processing module is used to post-process the enhanced speech signal to obtain the final speech signal.
[0008] Furthermore, the preprocessing module preprocesses the original speech signal, including:
[0009] Obtain a preset standard amplitude range, and adjust the amplitude of the original speech signal within the preset standard amplitude range based on automatic gain control;
[0010] High-pass filtering is applied to the calibrated original speech signal to remove the DC component from the original speech signal;
[0011] Bandpass filtering is applied to the original speech signal after high-pass filtering to remove high and low frequency noise from the original speech signal.
[0012] Furthermore, the separation module establishes a speech separation model, and separates the effective speech signal and noise signal from the preprocessed speech signal according to the speech separation model, including:
[0013] Collect speech sample data and noise sample data, and mix the speech audio in the speech sample data and the noise audio in the noise sample data to obtain mixed audio;
[0014] A training sample set is established based on the mixed audio and the corresponding speech tags of the mixed audio. A speech separation model is established and trained based on the training sample set to obtain the trained speech separation model.
[0015] The preprocessed speech signal is input into the trained speech separation model to obtain the effective speech signal and noise signal.
[0016] Furthermore, the step of establishing and training the speech separation model based on the training sample set includes:
[0017] The speech separation model is trained using a scale-invariant signal-to-noise ratio loss function.
[0018] Furthermore, the enhancement module performs anomaly filling on the effective speech signal according to the AI completion model to obtain an enhanced speech signal, including:
[0019] Identify the timbre features of valid speech signals, and vectorize the timbre features to obtain timbre feature vectors;
[0020] Collect the changes in timbre feature vectors between each audio frame of the valid speech signal, and identify abnormal speech segments based on the changes in timbre feature vectors;
[0021] Input the normal speech segment adjacent to the speech abnormal segment into the AI speech completion model to obtain the filled speech of the speech abnormal segment;
[0022] Speech filling is performed on all abnormal speech segments to obtain an enhanced speech signal.
[0023] Furthermore, the step of identifying abnormal speech segments based on changes in timbre feature vectors includes:
[0024] Calculate the vector change between each audio frame and its adjacent frames in the effective speech signal based on the changes in timbre feature vectors, and form a vector change sequence.
[0025] Establish a time window in the vector change sequence with any vector change as the center, calculate the average vector change within the time window, and determine the abnormal threshold of the corresponding vector change based on the average vector change within the time window.
[0026] The abnormal thresholds of all vector changes are statistically analyzed, and audio frames with vector changes greater than the corresponding abnormal thresholds are selected to obtain abnormal audio frames. Speech abnormal segments are determined based on all abnormal audio frames.
[0027] Furthermore, the step of determining the abnormal speech segment based on all abnormal audio frames includes:
[0028] The time interval of each abnormal audio frame is obtained, and abnormal audio frames with time intervals less than a preset time threshold are merged to obtain multiple initial abnormal segments.
[0029] The average value of the vector change of each initial abnormal segment is obtained. The initial abnormal segments are clustered according to the average value of the vector change, and multiple speech abnormal segments are obtained based on the clustering results.
[0030] Furthermore, the post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal, including:
[0031] The enhanced speech signal is subjected to moving average filtering to obtain the enhanced speech signal after moving average filtering.
[0032] Adaptive threshold filtering is applied to the enhanced speech signal after moving average filtering to obtain the final speech signal.
[0033] Furthermore, the post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal, and also includes:
[0034] The amplitude of the speech signal after adaptive threshold filtering is normalized to obtain the final speech signal.
[0035] On the other hand, the present invention also provides a method for noise reduction in narrowband satellite voice communication, comprising:
[0036] The raw speech signal from the satellite receiver is acquired, and the raw speech signal is preprocessed to obtain the preprocessed speech signal.
[0037] Establish a speech separation model, and separate the effective speech signal and noise signal from the preprocessed speech signal based on the speech separation model;
[0038] An AI completion model is established, and the effective speech signal is filled with anomalies based on the AI completion model to obtain an enhanced speech signal.
[0039] The enhanced speech signal is post-processed to obtain the final speech signal.
[0040] The narrowband satellite voice communication noise reduction system and method provided by this invention significantly improves the communication quality of voice signals and enhances communication reliability by performing multi-dimensional processing on narrowband satellite voice signals and performing voice separation and enhancement based on voice separation models and AI completion models. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 A schematic diagram of a narrowband satellite voice communication noise reduction system proposed in an embodiment of the present invention is shown;
[0043] Figure 2 The diagram shows the overall flowchart of a narrowband satellite voice communication noise reduction method proposed in an embodiment of the present invention. Detailed Implementation
[0044] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0045] This application provides a narrowband satellite voice communication noise reduction system, such as... Figure 1 As shown, it includes:
[0046] The preprocessing module acquires the raw speech signal from the satellite receiver, preprocesses it to obtain the preprocessed speech signal; the separation module establishes a speech separation model and separates the effective speech signal and noise signal from the preprocessed speech signal based on the speech separation model; the enhancement module establishes an AI completion model and fills in anomalies in the effective speech signal based on the AI completion model to obtain the enhanced speech signal; the post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal.
[0047] In this embodiment, the preprocessing module preprocesses the original speech signal, providing a standardized signal with stable amplitude and initial suppression of basic noise for subsequent modules. The preprocessed speech signal is then separated into effective speech signal and noise signal by the separation module. The effective speech signal is then anomaly filled by the enhancement module, which can effectively repair the speech loss problem caused by channel fading. Finally, the postprocessing module smooths out traces and optimizes the listening experience to obtain the final speech signal.
[0048] In some embodiments of this application, the preprocessing module preprocesses the original speech signal, including: obtaining a preset standard amplitude range, adjusting the amplitude of the original speech signal within the preset standard amplitude range based on automatic gain control; performing high-pass filtering on the calibrated original speech signal to filter out the DC component in the original speech signal; and performing band-pass filtering on the high-pass filtered original speech signal to filter out high and low frequency noise in the original speech signal.
[0049] In this embodiment, the signal amplitude is stabilized within the range of [-1, 1] by automatic gain control (AGC) to avoid processing errors caused by excessively strong or weak signals. A high-pass filter with a cutoff frequency of 100Hz is used to remove the DC component in the signal to avoid interference from DC offset on subsequent model inference. Finally, a bandpass filter is designed to filter out high-frequency and low-frequency noise that exceeds the transmission range of the narrowband satellite link, resulting in a preprocessed speech signal.
[0050] In some embodiments of this application, the separation module establishes a speech separation model and separates effective speech signals and noise signals from the preprocessed speech signal according to the speech separation model, including: collecting speech sample data and noise sample data; mixing the speech audio in the speech sample data and the noise audio in the noise sample data to obtain mixed audio; establishing a training sample set according to the mixed audio and the speech tags corresponding to the mixed audio; establishing and training the speech separation model according to the training sample set to obtain a trained speech separation model; and inputting the preprocessed speech signal into the trained speech separation model to obtain effective speech signals and noise signals.
[0051] In some embodiments of this application, the step of establishing and training a speech separation model based on a training sample set includes: training the speech separation model using a scale-invariant signal-to-noise ratio loss function.
[0052] In this embodiment, noise-free, clean speech samples of different genders, ages, and accents are collected as speech sample data. Typical satellite link noises, such as space electromagnetic noise, satellite equipment noise, multipath interference noise, and common environmental noises, such as wind and rain, are collected as noise sample data. The speech audios are manually labeled to obtain corresponding speech tags. The speech audios are randomly mixed with the noise audios to obtain mixed audio. The mixed audios and the corresponding speech tags are combined, and a speech separation model is established and trained using a scale-invariant signal-to-noise ratio loss function. The speech separation model then outputs the effective speech signal and noise signal corresponding to the preprocessed speech signal.
[0053] In some embodiments of this application, the enhancement module performs anomaly filling on the effective speech signal according to the AI completion model to obtain an enhanced speech signal, including: identifying the timbre features of the effective speech signal, vectorizing the timbre features to obtain timbre feature vectors; collecting the changes in timbre feature vectors between each audio frame of the effective speech signal, and identifying speech anomalous segments based on the changes in timbre feature vectors; inputting the normal speech segments adjacent to the speech anomalous segments into the AI speech completion model to obtain the filled speech for the speech anomalous segments; and performing speech filling on all speech anomalous segments to obtain an enhanced speech signal.
[0054] In this embodiment, the timbre features of the effective speech signal can be extracted based on the Mert model and vectorized. The abnormal speech segments are identified by the changes in the timbre feature vectors between audio frames. After removing the abnormal speech segments, the AI speech completion model outputs the fill speech based on the normal speech segments adjacent to the abnormal speech segments to complete the speech filling. In this embodiment, the length of the normal speech segments is equal to the length of the abnormal speech segments. By filling all the abnormal speech segments, an enhanced speech signal is obtained.
[0055] In some embodiments of this application, the step of identifying abnormal speech segments based on changes in timbre feature vectors includes: calculating the vector change between each audio frame and adjacent frames in the valid speech signal based on changes in timbre feature vectors, forming a vector change sequence; establishing a time window in the vector change sequence with any vector change as the center, calculating the average vector change within the time window, determining the abnormal threshold of the corresponding vector change based on the average vector change within the time window; statistically analyzing the abnormal thresholds of all vector changes, filtering out audio frames with vector changes greater than the corresponding abnormal thresholds to obtain abnormal audio frames, and determining abnormal speech segments based on all abnormal audio frames.
[0056] In this embodiment, the average value of the vector change of any audio frame and its two adjacent audio frames is calculated to obtain the vector change of the audio frame. The abnormal threshold of the audio frame is calculated based on the time window. In this embodiment, the abnormal adjustment parameter is preset and multiplied by the average vector change within the corresponding time window of the audio frame to obtain the abnormal threshold. Abnormal audio frames are selected and abnormal speech segments are determined based on the abnormal threshold.
[0057] In some embodiments of this application, determining the speech abnormal segment based on all abnormal audio frames includes: obtaining the time interval of each abnormal audio frame, merging abnormal audio frames with time intervals less than a preset time threshold to obtain multiple initial abnormal segments; obtaining the average value of the vector change of each initial abnormal segment, clustering each initial abnormal segment based on the average value of the vector change, and obtaining multiple speech abnormal segments based on the clustering results.
[0058] In this embodiment, abnormal audio frames with a time interval of less than 50ms are first merged to obtain initial abnormal segments. The K-means clustering algorithm is used to cluster each initial abnormal segment based on the magnitude of the average value of the vector change, thereby merging the initial abnormal segments in the same cluster. If there are normal audio frames between the merged initial abnormal segments in the valid speech signal, the corresponding normal audio frames are merged together to form a speech abnormal segment.
[0059] In some embodiments of this application, the post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal, including: performing moving average filtering on the enhanced speech signal to obtain the moving average filtered enhanced speech signal; and performing adaptive threshold filtering on the moving average filtered enhanced speech signal to obtain the final speech signal.
[0060] In this embodiment, a five-point moving average filter is used to smooth the enhanced speech signal, eliminating any transient pulse interference that may exist in the speech signal. At the same time, an adaptive threshold filter is designed to suppress residual weak noise a second time, while avoiding the filtering out of speech details.
[0061] In some embodiments of this application, the post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal, and further includes: normalizing the amplitude of the speech signal after adaptive threshold filtering to obtain the final speech signal.
[0062] In this embodiment, the amplitude of the final output voice signal is normalized to [-1, 1] to avoid distortion caused by signal overload and to ensure that it matches the output requirements of the satellite voice terminal equipment.
[0063] This invention also provides a method for noise reduction in narrowband satellite voice communication, such as... Figure 2 As shown, it includes:
[0064] S101: Acquire the raw speech signal from the satellite receiver, preprocess the raw speech signal, and obtain the preprocessed speech signal;
[0065] S102, Establish a speech separation model, and separate the effective speech signal and noise signal from the preprocessed speech signal according to the speech separation model;
[0066] S103, Establish an AI completion model, and fill in the abnormalities of the effective speech signal according to the AI completion model to obtain an enhanced speech signal;
[0067] S104, post-processes the enhanced speech signal to obtain the final speech signal.
[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0069] 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A narrowband satellite voice communication noise reduction system, characterized in that, include: The preprocessing module is used to acquire the raw speech signal from the satellite receiver, preprocess the raw speech signal, and obtain the preprocessed speech signal. The separation module is used to establish a speech separation model and separate the effective speech signal and noise signal from the preprocessed speech signal according to the speech separation model; The enhancement module is used to build an AI completion model, and fill in the abnormalities of the effective speech signal according to the AI completion model to obtain the enhanced speech signal. The post-processing module is used to post-process the enhanced speech signal to obtain the final speech signal; The enhancement module performs anomaly filling on the valid speech signal according to the AI completion model to obtain the enhanced speech signal, including: Identify the timbre features of valid speech signals, and vectorize the timbre features to obtain timbre feature vectors; Collect the changes in timbre feature vectors between each audio frame of the valid speech signal, and identify abnormal speech segments based on the changes in timbre feature vectors; Input the normal speech segment adjacent to the speech abnormal segment into the AI speech completion model to obtain the filled speech of the speech abnormal segment; Speech filling is performed on all abnormal speech segments to obtain an enhanced speech signal; The method of identifying abnormal speech segments based on changes in timbre feature vectors includes: Calculate the vector change between each audio frame and its adjacent frames in the effective speech signal based on the changes in timbre feature vectors, and form a vector change sequence. Establish a time window in the vector change sequence with any vector change as the center, calculate the average vector change within the time window, and determine the abnormal threshold of the corresponding vector change based on the average vector change within the time window. Calculate the abnormal threshold of all vector changes, filter out audio frames whose vector changes are greater than the corresponding abnormal threshold, obtain abnormal audio frames, and determine the abnormal speech segments based on all abnormal audio frames. The step of determining abnormal speech segments based on all abnormal audio frames includes: The time interval of each abnormal audio frame is obtained, and abnormal audio frames with time intervals less than a preset time threshold are merged to obtain multiple initial abnormal segments. The average value of the vector change of each initial abnormal segment is obtained. The initial abnormal segments are clustered according to the average value of the vector change, and multiple speech abnormal segments are obtained based on the clustering results.
2. The narrowband satellite voice communication noise reduction system according to claim 1, characterized in that, The preprocessing module preprocesses the original speech signal, including: Obtain a preset standard amplitude range, and adjust the amplitude of the original speech signal within the preset standard amplitude range based on automatic gain control; High-pass filtering is applied to the calibrated original speech signal to remove the DC component from the original speech signal; Bandpass filtering is applied to the original speech signal after high-pass filtering to remove high and low frequency noise from the original speech signal.
3. The narrowband satellite voice communication noise reduction system according to claim 1, characterized in that, The separation module establishes a speech separation model, and separates the effective speech signal and noise signal from the preprocessed speech signal according to the speech separation model, including: Collect speech sample data and noise sample data, and mix the speech audio in the speech sample data and the noise audio in the noise sample data to obtain mixed audio; A training sample set is established based on the mixed audio and the corresponding speech tags of the mixed audio. A speech separation model is established and trained based on the training sample set to obtain the trained speech separation model. The preprocessed speech signal is input into the trained speech separation model to obtain the effective speech signal and noise signal.
4. The narrowband satellite voice communication noise reduction system according to claim 3, characterized in that, The process of establishing and training a speech separation model based on a training sample set includes: The speech separation model is trained using a scale-invariant signal-to-noise ratio loss function.
5. The narrowband satellite voice communication noise reduction system according to claim 1, characterized in that, The post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal, including: The enhanced speech signal is subjected to moving average filtering to obtain the enhanced speech signal after moving average filtering. Adaptive threshold filtering is applied to the enhanced speech signal after moving average filtering to obtain the final speech signal.
6. The narrowband satellite voice communication noise reduction system according to claim 5, characterized in that, The post-processing module performs post-processing on the enhanced speech signal to obtain the final speech signal, and also includes: The amplitude of the speech signal after adaptive threshold filtering is normalized to obtain the final speech signal.
7. A method for noise reduction in narrowband satellite voice communication, characterized in that, include: The raw speech signal from the satellite receiver is acquired, and the raw speech signal is preprocessed to obtain the preprocessed speech signal. Establish a speech separation model, and separate the effective speech signal and noise signal from the preprocessed speech signal based on the speech separation model; An AI completion model is established, and the effective speech signal is filled with anomalies based on the AI completion model to obtain an enhanced speech signal. The enhanced speech signal is post-processed to obtain the final speech signal; The process of filling in anomalies in the effective speech signal using an AI completion model to obtain an enhanced speech signal includes: Identify the timbre features of valid speech signals, and vectorize the timbre features to obtain timbre feature vectors; Collect the changes in timbre feature vectors between each audio frame of the valid speech signal, and identify abnormal speech segments based on the changes in timbre feature vectors; Input the normal speech segment adjacent to the speech abnormal segment into the AI speech completion model to obtain the filled speech of the speech abnormal segment; Speech filling is performed on all abnormal speech segments to obtain an enhanced speech signal; The method of identifying abnormal speech segments based on changes in timbre feature vectors includes: Calculate the vector change between each audio frame and its adjacent frames in the effective speech signal based on the changes in timbre feature vectors, and form a vector change sequence. Establish a time window in the vector change sequence with any vector change as the center, calculate the average vector change within the time window, and determine the abnormal threshold of the corresponding vector change based on the average vector change within the time window. Calculate the abnormal threshold of all vector changes, filter out audio frames whose vector changes are greater than the corresponding abnormal threshold, obtain abnormal audio frames, and determine the abnormal speech segments based on all abnormal audio frames. The step of determining abnormal speech segments based on all abnormal audio frames includes: The time interval of each abnormal audio frame is obtained, and abnormal audio frames with time intervals less than a preset time threshold are merged to obtain multiple initial abnormal segments. The average value of the vector change of each initial abnormal segment is obtained. The initial abnormal segments are clustered according to the average value of the vector change, and multiple speech abnormal segments are obtained based on the clustering results.