A classification-based sibilance suppression method

Abstract

The present application belongs to the technical field of digital audio processing, and provides a classification-based sibilance suppression method to solve the problem that the prior art causes perceptible loss of useful speech information when suppressing sibilance due to the lack of fine-grained distinction between sibilance and useful speech. The present application includes two stages of classifier construction and sibilance suppression. The spectrum of a training audio frame is extracted, the spectral amplitude data of the sibilance frequency band is grouped, the sum of the spectral amplitudes in each group is calculated, and the amplitudes of each group are spliced to form a feature, on the basis of which a sibilance classifier is trained. In sibilance suppression, the sibilance type probability output by the sibilance classifier is used to perform fine-grained type prediction on the audio data to be processed in units of frames, and the prediction results of consecutive frames are smoothed through median filtering. Finally, the relative energy of sibilance and non-sibilance is calculated to attenuate the sibilance component, thereby effectively suppressing the sibilance in the audio while maintaining the original speech information.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital audio processing, and specifically provides a classification-based sibilant suppression method. Background Art

[0002] Sibilants are the sounds produced when the tip of the tongue touches the upper front teeth during human pronunciation; usually, when people read words such as "zi", "ci", "si", etc., sibilants are likely to occur. Because the frequency of sibilants is relatively high and they sound relatively sharp, sibilants in audio may affect the listening experience of the audience; after professional recording of songs or dubbing, it is usually necessary to tune the recorded audio to improve the listening effect, and suppressing sibilants in the audio is one of the steps.

[0003] Sibilant suppression can be achieved through frequency filtering (such as an EQ equalizer), and some audio processing software (such as Adobe Audition) has the function of sibilant suppression, and its technology has not been made public. At present, there are few reports on sibilant suppression methods in the literature; in the disclosed technologies, the invention patent "202110469577.7, Audio Signal Processing Method and Device" attenuates sibilants by performing dynamic compression processing on the sibilant frequency band. Since the sibilant frequency band also contains useful speech components, and this method uniformly processes the audio containing speech and sibilants without distinction, it will cause loss of speech information. The target of the invention patent "202210225928.4, A Dubbing Audio Analysis Method, Device, Electronic Device and Storage Medium" is to improve the quality of audio containing abnormal sounds (such as saliva sounds, sibilants, bubble sounds, and breathing sounds, etc.). According to the spectrogram image of the currently to-be-processed audio and the spectrogram images of pre-classified problematic audios, image recognition technology and algorithm clustering analysis technology are used to judge whether there is an abnormal pronunciation in the dubbing audio and determine the type of abnormality. Finally, the dubbing audio is corrected by correcting the frequency, amplitude, timbre, pitch, and loudness of the sound; this method uses image recognition technology to identify various types such as saliva sounds, sibilants, bubble sounds, and breathing sounds. Since the spectrogram image features of these abnormal types of audio have similar components, the classification accuracy is not good enough; in addition, the same as the previous method, this method also corrects the entire audio, inevitably damaging the original speech information.

[0004] In summary, the current sibilant suppression methods do not make fine-grained distinctions between sibilants and useful speech, resulting in perceivable loss of useful speech information while suppressing sibilants. Summary of the Invention

[0005] The purpose of the present invention is to provide a more effective classification-based sibilant suppression method to solve the problem that the existing sibilant suppression methods do not make fine-grained distinctions between sibilants and useful speech, resulting in perceivable loss of useful speech information when suppressing sibilants.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A classification-based method for suppressing dental sounds, characterized by comprising two stages: classifier construction and dental sound suppression;

[0008] The classifier construction includes the following steps:

[0009] A1. Generate training samples:

[0010] A1-1. Initialize the training feature matrix R and training label vector L to be empty;

[0011] A1-2. Generate a training feature matrix R and a training label vector L based on the training audio set. The training audio set contains training audios that are previously labeled as either sibilant or non-sibilant. For each training audio, perform the following:

[0012] A1-2-1. Divide the audio data into audio frames of fixed length, with overlap between adjacent frames, to obtain I audio frames; and perform windowing processing on the data of each audio frame, followed by short-time Fourier transform to obtain its corresponding spectrum;

[0013] A1-2-2. Suppose there are N+1 data points corresponding to each time step in the spectral amplitude data, and the preset characteristic frequency range is [f1, f2]. Then calculate the characteristic frequency points respectively. and

[0014]

[0015]

[0016] Where f1 and f2 are the lowest and highest characteristic frequencies, respectively. This is a floor function;

[0017] A1-2-3. For each audio frame, perform the following:

[0018] A1-2-3-1. For the i-th audio frame, 1≤i≤i, from arrive Group the spectral amplitude data corresponding to every M frequency points into a group. Stop grouping when the number of remaining frequency points is less than M, where M is the preset grouping length. Calculate the sum of the amplitude data corresponding to all frequency points in each group, and arrange the summation results of each group in sequence to form the training feature vector r. i ;

[0019] A1-2-3-2. The training feature vector r i Add the columns to the training feature matrix R;

[0020] A1-2-3-3. Let l i Let l be the type label for the i-th audio frame. If the audio to which the current audio frame belongs is sibilant audio, then set l. i =1; otherwise, set l i =0;

[0021] A1-2-3-4. Type label l i Add to the training label vector L;

[0022] A2. Training the dental consonant classifier: The classifier Q is trained using the training feature matrix R and the training label vector L. This classifier is a binary classifier, and its input is the training feature vector and its output is the probability that the training feature vector is predicted to be a dental consonant.

[0023] The sibilance suppression includes the following steps:

[0024] B1. Generate test samples:

[0025] B1-1. Initialize the test feature matrix R′ to be empty;

[0026] B1-2. Generate the test feature matrix R′, specifically as follows:

[0027] B1-2-1. Divide the audio data to be processed into audio frames of fixed length, with overlap between adjacent frames, resulting in I′ audio frames; perform windowing processing on the data of each audio frame, and then perform a short-time Fourier transform to obtain its corresponding spectrum, where the spectral amplitude data of the i-th audio frame is c. i ;

[0028] B1-2-2. Perform the following for each audio frame:

[0029] B1-2-2-1. For the i-th audio frame, 1≤i≤i′, from arrive Group the spectral amplitude data corresponding to every M frequency points into a group. Stop grouping when the number of remaining frequency points is less than M. Calculate the sum of the amplitude data corresponding to all frequency points in each group, and arrange the summation results of each group in sequence to form the test feature vector r. i ′;

[0030] B1-2-2-2. r i 'Append column by column to the test feature matrix R';

[0031] B2. Audio Frame Classification:

[0032] B2-1. Define an array K of velar consonant types with length I′;

[0033] B2-2. Perform the following for each audio frame:

[0034] B2-2-1. The i-th column eigenvector r of R′ i The input is fed into classifier Q to obtain the probability p that the feature vector is predicted to be a dental consonant. i ;

[0035] B2-2-2. Assign a value to the i-th element K[i] of the dental consonant type array K:

[0036]

[0037] Where T0 and T2 are the set credible non-acidic sound probability threshold and credible acidic sound probability threshold, respectively;

[0038] B2-3. Perform median filtering on the data in the dental consonant type array K;

[0039] B3. Attenuation of sibilance:

[0040] B3-1. Initialize i = 1, reference energy Y0 = 1;

[0041] B3-2. Calculate the eigenvector r i The sum of the values ​​of all elements of ' is denoted as Y. i ;

[0042] B3-3. If K[i] = 1, then update Y0 to Y i Otherwise, proceed to step B3-4;

[0043] B3-4. If K[i] = 2, then update the spectral amplitude data vector c. i The The first to the second One element:

[0044]

[0045] Among them, c i [j] is c i The j-th element;

[0046] Otherwise, proceed to step B3-5;

[0047] B3-5. Update i to i+1, and determine if i is greater than I′. If yes, go to step B3-6; otherwise, go to step B3-2.

[0048] B3-6. Perform a short-time inverse Fourier transform using the updated spectral data to obtain the processed audio.

[0049] Furthermore, the minimum characteristic frequency f1 has a value range of 4000≤f1≤5000.

[0050] Furthermore, the range of the highest characteristic frequency f2 is: 10000≤f2≤12000.

[0051] Furthermore, the grouping length M has the following range: 1 ≤ M ≤ 10.

[0052] Furthermore, the range of the credible non-abrasive probability threshold T0 is: 0 < T0 < 0.5.

[0053] Furthermore, the range of the reliable dental sound probability threshold T2 is: 0.5 < T2 < 1.

[0054] Based on the above technical solution, the beneficial effects of the present invention are as follows:

[0055] This invention provides a classification-based sibilance suppression method. The method extracts the spectrum of training audio frames, groups the spectral amplitude data of sibilance frequency bands, calculates the sum of the spectral amplitudes within each group, and then concatenates the amplitude sums of each group to form a feature. Based on this feature, a sibilance classifier is trained. In sibilance suppression, based on the sibilance type probability output by the sibilance classifier, fine-grained type prediction is performed on the audio data to be processed, frame by frame. Median filtering is used to smooth the prediction results of consecutive frames. Finally, by calculating the relative energy of sibilances and non-sibilances, the sibilance component is attenuated, thereby effectively suppressing sibilances in the audio while preserving the original speech information. Attached image description:

[0056] Figure 1 This is a flowchart illustrating the classification-based sibilance suppression method of the present invention, which includes two stages: classifier construction and sibilance suppression.

[0057] Figure 2 The above are comparison diagrams of the effects of classification-based sibilance suppression methods in the embodiments of the present invention, wherein (a) is the spectrum amplitude diagram of the test audio, (b) is the spectrum amplitude diagram after sibilance removal using the audio processing software Adobe Audition, and (c) is the spectrum amplitude diagram after processing by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0059] This embodiment provides a classification-based method for suppressing dental sounds, the process of which is as follows: Figure 1 As shown, this method consists of two stages: classifier construction and sibilance suppression.

[0060] The following section will provide a more detailed explanation of the classifier construction steps:

[0061] A1. Generate training samples:

[0062] A1-1. Initialize the training feature matrix R and training label vector L to be empty;

[0063] A1-2. Generating the training feature matrix R and training label vector L: Given a set of training audio files, where each audio file is labeled as either sibilant or non-sibilant, perform the following steps for each audio file in the training audio set:

[0064] In this embodiment, there are 40 training audios in total, including 20 sibilant audios and 20 non-sibilant audios. Each audio is a mono audio with a sampling rate of 48000Hz. The following steps are explained using the processing of the first sibilant audio as an example.

[0065] A1-2-1. Divide the audio data into audio frames of fixed length, with overlap between adjacent frames. Let the total number of audio frames obtained be I. Window the audio frame data and then perform a short-time Fourier transform to obtain its corresponding spectrum.

[0066] In this embodiment, the audio data is divided into audio frames with a length of 512 sampling points. There is a half-frame overlap between adjacent frames, and the total number of audio frames obtained is I = 10. The windowing process uses Hamming windowing.

[0067] A1-2-2. Suppose that in the spectral amplitude data of an audio frame, there are N+1 amplitude data points at each time point, corresponding to frequencies from 0 to the highest frequency f of the signal. max Given N+1 frequency points, and a preset characteristic frequency range of [f1, f2], where f1 and f2 are the lowest and highest frequencies within the characteristic frequency range, respectively, the corresponding characteristic frequency points are calculated as follows:

[0068]

[0069]

[0070] in, and These are the frequency point numbers corresponding to f1 and f2, respectively. This is a floor function;

[0071] In this embodiment, the lowest characteristic frequency f1 ranges from 4000 to 5000, where f1 is 4000; the highest characteristic frequency f2 ranges from 10000 to 12000, where f2 is 10000. In the amplitude data of the spectrum, each moment of an audio frame contains N+1 = 257 amplitude data points, corresponding to frequencies from 0 to the highest signal frequency f. maxThe frequency point number corresponding to f1 = 4000Hz is calculated from the 257 frequency points between 24000Hz using the formula. The frequency point number corresponding to f2 = 10000Hz

[0072] A1-2-3. For each i satisfying 1≤i≤I, perform:

[0073] In this embodiment, I = 10, therefore for each i satisfying 1 ≤ i ≤ 10, the following is performed:

[0074] A1-2-3-1. For the i-th frame of audio, from... arrive Group the spectral amplitude data corresponding to every M frequency points into a group. Stop grouping when the number of remaining frequency points is less than M, where M is the preset grouping length. For each group, calculate the sum of the amplitude values ​​corresponding to all frequency points in the group. Arrange the summation results of each group into a vector column, denoted as the training feature vector r of the i-th frame. i ;

[0075] In this embodiment, the group length M ranges from 1 to M to 10, where M is 2; taking the i=1th frame as an example, starting from the frequency point number n... f2 =107 to frequency point number The spectral amplitude data corresponding to every M=2 frequency points are grouped into 32 groups. The sum of the amplitude values ​​corresponding to the frequency points in each group is calculated as follows: 0.432, 1.842, ..., 0.592. These 32 summed data are arranged into a vector to obtain the training feature vector r1 = [0.432, 1.842, ..., 0.592] for the first frame of the current audio. T ;

[0076] A1-2-3-2. r i The features are appended column-wise to the training feature matrix R, i.e., R = [R, r i ];

[0077] A1-2-3-3. Let l i Let l be the type label for the i-th frame. If the audio of the current frame is sibilant audio, then set l. i =1, otherwise l i =0;

[0078] In this embodiment, taking the i=1th frame as an example, since the audio of the current frame is sibilant audio, then set l i =1;

[0079] A1-2-3-4. l i Append it to the training label vector L, i.e., L = [L, l i];

[0080] In this embodiment, taking the i=1 frame as an example, l1 is appended to the training label vector L, that is, L=[L,l1];

[0081] A2. Training the dental consonant classifier: Train a binary classifier Q using the training feature matrix R and the training label vector L. The classifier takes a column of feature vectors and outputs the probability that the feature vector is predicted to be a dental consonant.

[0082] In this embodiment, an SVM classifier Q is trained using training data R and L, wherein the SVM parameters are set as follows: the penalty coefficient C for erroneous terms is set to 1, and the kernel function is set to a linear function.

[0083] The following is a further detailed explanation of the sibilance suppression steps:

[0084] B1. Generate test samples:

[0085] B1-1. Initialize the test feature matrix R′ to be empty;

[0086] B1-2. Generate the test feature matrix R′:

[0087] In this embodiment, the test audio is a mono audio with a sampling rate of 48000Hz and a duration of 5s, which includes 3 sibilance intervals. The first sibilance appears at 1.428s and lasts for 0.069s, the second sibilance appears at 2.121s and lasts for 0.027s, and the third sibilance appears at 2.324s and lasts for 0.117s.

[0088] B1-2-1. Divide the input test audio data into audio frames of fixed length, with overlap between adjacent frames. Let the total number of audio frames be I′. Window the data of each audio frame and then perform a short-time Fourier transform to obtain its corresponding spectrum. Let c be the vector formed by the spectrum amplitude data of the i-th frame. i ;

[0089] In this embodiment, the audio data is divided into audio frames with a length of 512 sampling points, and adjacent frames overlap by half a frame length, resulting in a total number of audio frames I′ = 938. Frames 268 to 280 constitute the first sibilance interval, frames 398 to 402 constitute the second sibilance interval, and frames 436 to 457 constitute the third sibilance interval. A Hamming window is used for windowing, and after performing a short-time Fourier transform, the spectrum corresponding to each audio frame is obtained, and its amplitude diagram is shown below. Figure 2 As shown in (a), taking the i=268th frame as an example, the vector c formed by the spectral amplitude data of this frame... 268 =[0.252,0.171,……,0.646]T ;

[0090] B1-2-2. For each i satisfying 1≤i≤I', perform:

[0091] In this embodiment, I′=938, therefore for each i satisfying 1≤i≤938, we perform:

[0092] B1-2-2-1. For the i-th frame of audio, from... arrive Group the spectral amplitude data corresponding to every M frequency points into a group. Stop grouping when the number of remaining frequency points is less than M. For each group, calculate the sum of the amplitude values ​​corresponding to all frequency points in the group. Arrange the summation results of each group into a vector, denoted as the test feature vector r of the i-th frame. i ′;

[0093] In this embodiment, taking the i=268th frame as an example, the frequency point numbering... Frequency point number The spectral amplitude data corresponding to every M=2 frequency points are grouped into 32 groups. The sum of the amplitude values ​​corresponding to the frequency points in each group is calculated as follows: 0.662, 0.546, ..., 1.127. These 32 summed data are arranged into a vector to obtain the test feature vector r2′ of the 268th frame of the current audio. 68 = [0.662, 0.546, ..., 1.127] T ;

[0094] B1-2-2-2. r i The columns are appended to the test feature matrix R′, i.e., R′ = [R′, r i ′];

[0095] B2. Audio Frame Classification:

[0096] B2-1. Define an array K of velar consonant types with length I′;

[0097] B2-2. For each i satisfying 1≤i≤I′, perform:

[0098] In this embodiment, I′=938, therefore for each i satisfying 1≤i≤938, we perform:

[0099] B2-2-1. The i-th column eigenvector r of R′ i The input is fed into classifier Q to obtain the probability p that the feature vector is predicted to be a dental consonant. i ;

[0100] In this embodiment, taking the i=268th column as an example, the feature vector r2′ 68The input is fed into classifier Q to obtain the probability p that the feature vector is predicted to be a dental consonant. 268 =0.815;

[0101] B2-2-2. Assign a value to the i-th element K[i] of the dental consonant type array K:

[0102]

[0103] Where T0 and T2 are the set credible non-acidic sound probability threshold and credible acidic sound probability threshold, respectively;

[0104] In this embodiment, the range of the credible non-abrasive probability threshold T0 is 0 < T0 < 0.5, where T0 is 0.3; the range of the credible abrasive probability threshold T2 is 0.5 < T2 < 1, where T2 is 0.7; taking the i = 268th element as an example, since p 268 =0.815, p 268 ≥T2, thus K

[268] =2;

[0105] B2-3. Perform median filtering on the data in the dental consonant type array K;

[0106] In this embodiment, after step B2-2, each element of the dental consonant type array K is assigned a value. For example, the 12 element values ​​between the 230th and 241st elements of K are: 0,0,0,0,2,1,1,1,0,1,0,0. Step B2-3 performs median filtering on the data of K to obtain updated data. For example, the 12 element values ​​between the 230th and 241st elements of K are updated to: 0,0,0,0,1,1,1,1,1,0,0,0. The median filtering length is set to 5.

[0107] B3. Attenuation of sibilance:

[0108] B3-1. Initialize i = 1, reference energy Y0 = 1;

[0109] B3-2. Calculate r i The sum of the values ​​of all elements of ' is denoted as Y. i ;

[0110] In this embodiment, when i = 238 frames, r2′ is calculated. 38 The sum of the values ​​of all elements yields Y. 238 =20.950;

[0111] B3-3. If K[i] = 1, then update Y0 to Y i Otherwise, proceed to step B3-4;

[0112] In this embodiment, when i = 238 frames, since K

[238] = 1, Y0 is updated to Y. 238 That is, Y0 = 20.950;

[0113] B3-4. If K[i] = 2, then update the spectral amplitude data vector c. i The The first to the second One element:

[0114]

[0115] Among them, c i [j] is c i The j-th element;

[0116] Otherwise, proceed to step B3-5;

[0117] In this embodiment, when i = 268 frames, since K

[268] = 2, the spectral amplitude data vector c is updated. 268 The The first to the second One element:

[0118]

[0119] Before the update, c 268 The 43rd to 107th elements are 0.317, 0.367, ..., 0.357 respectively. After step B3-4, these elements are updated to: 0.112, 0.130, ..., 0.126 respectively.

[0120] B3-5. Update i to i+1, and determine if i is greater than I′. If yes, go to step B3-6; otherwise, go to step B3-2.

[0121] In this embodiment, I′=938, so i is updated to i+1, and it is determined whether i is greater than 938. If yes, proceed to step B3-6; otherwise, proceed to step B3-2.

[0122] B3-6. Perform a short-time inverse Fourier transform using the updated spectral data to obtain the processed audio.

[0123] The following comparison, with reference to an embodiment, shows the sibilance suppression effect of the method of the present invention with that of the audio processing software Adobe Audition, and the results are as follows. Figure 2 As shown; from Figure 2 As can be seen in (a), sibilance appears at 1.428s, 2.121s, and 2.324s in the original audio, with the corresponding spectral amplitude graph showing higher highlights in the 4K to 10K frequency range; after processing by Adobe Audition and this embodiment, as shown... Figure 2 As shown in (b) and (c), the highlights are reduced, indicating that the sibilant components are suppressed. The suppression effect of the method of the present invention is more obvious than that of Adobe Audition, and there is no loss of speech content perceived subjectively.

[0124] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A classification-based method for suppressing dental consonants, characterized in that, It includes two stages: classifier construction and sibilance suppression; The classifier construction includes the following steps: A1. Generate training samples: A1-1. Initialize the training feature matrix R and training label vector L to be empty; A1-2. Generate a training feature matrix R and a training label vector L based on the training audio set. The training audio set contains training audios that are previously labeled as either sibilant or non-sibilant. For each training audio, perform the following: A1-2-1. Divide the audio data into audio frames of fixed length, with overlap between adjacent frames, to obtain I audio frames; and perform windowing processing on the data of each audio frame, followed by short-time Fourier transform to obtain its corresponding spectrum; A1-2-2. Suppose there are N+1 data points corresponding to each time step in the spectral amplitude data, and the preset characteristic frequency range is [f1, f2]. Then calculate the characteristic frequency points respectively. and Where f1 and f2 are the lowest and highest characteristic frequencies, respectively. This is a floor function; A1-2-3. For each audio frame, perform the following: A1-2-3-1. For the i-th audio frame, 1≤i≤i, from arrive Group the spectral amplitude data corresponding to every M frequency points into a group. Stop grouping when the number of remaining frequency points is less than M, where M is the preset grouping length. Calculate the sum of the amplitude data corresponding to all frequency points in each group, and arrange the summation results of each group in sequence to form the training feature vector r. i ; A1-2-3-2. The training feature vector r i Add the columns to the training feature matrix R; A1-2-3-3. Let l i Let l be the type label for the i-th audio frame. If the audio to which the current audio frame belongs is sibilant audio, then set l. i =1; otherwise, set l i =0; A1-2-3-4. Type label l i Add to the training label vector L; A2. Training the dental consonant classifier: The classifier Q is trained using the training feature matrix R and the training label vector L. This classifier is a binary classifier, and its input is the training feature vector and its output is the probability that the training feature vector is predicted to be a dental consonant. The sibilance suppression includes the following steps: B1. Generate test samples: B1-1. Initialize the test feature matrix R′ to be empty; B1-2. Generate the test feature matrix R′, specifically as follows: B1-2-1. Divide the audio data to be processed into audio frames of fixed length, with overlap between adjacent frames, resulting in I′ audio frames; perform windowing processing on the data of each audio frame, and then perform a short-time Fourier transform to obtain its corresponding spectrum, where the spectral amplitude data of the i-th audio frame is c. i ; B1-2-2. Perform the following for each audio frame: B1-2-2-1. For the i-th audio frame, 1≤i≤i′, from arrive Group the spectral amplitude data corresponding to every M frequency points into a group. Stop grouping when the number of remaining frequency points is less than M. Calculate the sum of the amplitude data corresponding to all frequency points in each group, and arrange the summation results of each group in sequence to form the test feature vector r. i ′; B1-2-2-2. r i 'Append column by column to the test feature matrix R'; B2. Audio Frame Classification: B2-1. Define an array K of velar consonant types with length I′; B2-2. Perform the following for each audio frame: B2-2-1. The i-th column eigenvector r of R′ i The input is fed into classifier Q to obtain the probability p that the feature vector is predicted to be a dental consonant. i ; B2-2-2. Assign a value to the i-th element K[i] of the dental consonant type array K: Where T0 and T2 are the set credible non-acidic sound probability threshold and credible acidic sound probability threshold, respectively; B2-3. Perform median filtering on the data in the dental consonant type array K; B3. Attenuation of sibilance: B3-1. Initialize i = 1, reference energy Y0 = 1; B3-2. Calculate the eigenvector r i The sum of the values ​​of all elements of ' is denoted as Y. i ; B3-3. If K[i] = 1, then update Y0 to Y i Otherwise, proceed to step B3-4; B3-4. If K[i] = 2, then update the spectral amplitude data vector c. i The The first to the second One element: Among them, c i [j] is c i The j-th element; Otherwise, proceed to step B3-5; B3-5. Update i to i+1, and determine if i is greater than I′. If yes, go to step B3-6; otherwise, go to step B3-2. B3-6. Perform a short-time inverse Fourier transform using the updated spectral data to obtain the processed audio.

2. The classification-based sibilance suppression method according to claim 1, characterized in that, The minimum characteristic frequency f1 has a range of values: 4000≤f1≤5000.

3. The classification-based sibilance suppression method according to claim 1, characterized in that, The range of the highest characteristic frequency f2 is: 10000≤f2≤12000.

4. The classification-based sibilance suppression method according to claim 1, characterized in that, The grouping length M has the following value range: 1≤M≤10.

5. The classification-based sibilance suppression method according to claim 1, characterized in that, The range of the credible non-aesque probability threshold T0 is: 0 < T0 < 0.

5.

6. The classification-based sibilance suppression method according to claim 1, characterized in that, The range of the reliable sibilance probability threshold T2 is: 0.5 < T2 < 1.

Images