A beautiful voice filter matching method

Abstract

This invention discloses a method for matching vocal enhancement filters, comprising: acquiring target audio; importing the target audio into a pre-constructed filter analysis model to obtain a target filter; and using the target filter to perform reverberation on the target audio to generate a target reverberated audio. The technical advantages of this invention are: through automated filter selection, it selects and applies filters based on the user's voice, without relying on the experience of a sound engineer, achieving a one-click enhancement effect.

Description

Technical Field

[0001] The invention relates to the field of audio and video processing, and in particular to a method for matching a beautiful voice filter. Background Art

[0002] The existing karaoke effects allow users to choose scenes. A scene is a combination of effect parameters set based on experience. There are also parameters open to users for adjustment. However, what can be switched are the effect processing process, not the final effect target. However, on the one hand, the preset parameters are manually adjusted based on the experience of the tuner. The style and aesthetics are closely related to people, and the direction tends to be fixed. On the other hand, users switch the effect processing process, not the final effect target. In order to achieve the final effect, different sounds require different processing. Summary of the invention

[0003] In order to solve the technical problem in the prior art that reverberation adjustment cannot be performed without the experience of the tuner, the present invention provides a method for matching a beautiful voice filter, comprising:

[0004] Get the target audio;

[0005] Importing the target audio into a pre-built filter analysis model to obtain a target filter;

[0006] The target audio is reverberated using a target filter to generate a target reverberated audio.

[0007] This technical solution, through automatic filter selection, selects and applies filters for the user's voice, does not rely on the tuner's experience, and achieves a one-click beautification effect;

[0008] Preferably, the method further comprises: constructing the filter analysis model, specifically comprising:

[0009] Use sound source separation tools to extract reverberant vocal data from original sound data;

[0010] Performing reverberation separation on the reverberation vocal data to obtain pure vocal data and reverberation data;

[0011] Classification training is performed on the pure vocal data and the reverberation data to obtain a plurality of singer-vocal statistical results and a plurality of song-reverberation statistical results; the singer-vocal statistical results include the frequency spectrum distribution, dynamic differential parameter distribution, and effect sound volume of the target singer; the song-reverberation statistical results include the reverberation frequency spectrum distribution, reverberation time, and reverberation volume of the target song;

[0012] A cluster analysis is performed based on the multiple singer-vocal statistics and the multiple song-reverberation statistics to generate a filter analysis model.

[0013] Further preferably, the reverberation separation specifically includes:

[0014] Use linear predictive coding to perform reverberation separation to obtain pure vocal data and reverberation data;

[0015] The cost function of the linear predictive coding is Wherein, R is an impulse response filter, O is the order of the impulse response filter, and N is the length of the impulse response filter.

[0016] Further preferably, the length of the impulse response filter of the linear prediction coding is 0.5 seconds.

[0017] Further preferably, the classification training specifically includes:

[0018] Dividing the pure vocal data into pure vocal data sets according to a preset vocal block length;

[0019] Dividing the reverberation data into reverberation data sets according to a preset reverberation block length;

[0020] According to a preset vocal statistical algorithm, the pure vocal data set is counted to generate a singer-vocal statistical result;

[0021] According to a preset reverberation statistics algorithm, the reverberation data set is counted to generate a song-reverberation statistics result.

[0022] Further preferably, the length of the vocal block is 0.5 seconds;

[0023] The reverberation block length is 0.5 seconds.

[0024] Further preferably, the human voice statistical algorithm specifically includes:

[0025] Get all vocal chunks of the target singer;

[0026] Obtaining the vocal blocks in which the target singer is currently speaking from among all the vocal blocks, to form a target vocal statistics set;

[0027] The MFCC coefficients, the average volume of each block, and the average volume of the wet sound part of the target vocal statistical set are counted to generate singer-vocal statistical results;

[0028] The reverberation statistics algorithm specifically includes:

[0029] Obtain all reverberation blocks of the target song to form a target reverberation statistics set;

[0030] The MFCC coefficients, the average volume of each block, and the average volume of the wet sound part of the target reverberation statistical set are statistically analyzed to generate song-reverberation statistical results.

[0031] Further preferably, the cluster analysis specifically includes:

[0032] Specifically, the method includes: importing the plurality of singer-voice statistics and the song-reverberation statistics into a K-means algorithm to generate a filter feature vector clustering set;

[0033] Selecting a preset number of clusters, from among the filter feature vector clusters, a filter feature vector cluster with the largest number of feature vectors to generate a filter analysis model;

[0034] Further preferably, the K value of the K-means algorithm is 2 times the preset number of clusters to be selected.

[0035] Further preferably, importing the target audio into a pre-built filter analysis model specifically includes:

[0036] Divide the target audio into target audio block sets according to a preset target audio block length;

[0037] Arrange the target audio block set in chronological order, and make the MFCC coefficient of the current audio block equal to the MFCC coefficient of the first target audio block in the target audio block set;

[0038] The MFCC coefficients of the target audio block are counted sequentially to update the MFCC coefficients of the current audio block; when the MFCC coefficients of the current audio block are updated, the change amount of the MFCC coefficients of the current audio block is not higher than 1 decibel.

[0039] The present invention includes at least one of the following technical effects:

[0040] (1) Through automated filter selection, filters are selected and applied based on the user's voice, without relying on the tuner's experience, achieving a one-click beautification effect;

[0041] (2) By limiting the update and change of MFCC coefficients, stable MFCC coefficient statistics are achieved, reducing the interference of noise and fluctuations on filter selection, and improving the reliability and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0043] Figure 1This is a schematic diagram of the process of Example 1 of the present invention;

[0044] Figure 2 This is a schematic diagram of the process of Embodiment 2 of the present invention;

[0045] Figure 3 This is a schematic diagram of the process of Embodiment 3 of the present invention;

[0046] Figure 4 This is a schematic diagram of the process of Example 4 of the present invention;

[0047] Figure 5 This is a schematic diagram of the process of Example 5 of the present invention;

[0048] Figure 6 This is a schematic diagram of the process of Example 6 of the present invention;

[0049] Figure 7 It is a schematic diagram of the process of linear predictive coding of the present invention. DETAILED DESCRIPTION

[0050] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0051] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections.

[0052] In order to simplify the drawings, only the parts related to the present invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to simplify the drawings and facilitate understanding, in some figures, only one of the parts with the same structure or function is schematically drawn or marked. In this article, "one" not only means "only one", but also means "more than one".

[0053] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0054] In addition, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0056] Embodiment 1:

[0057] like Figure 1 As shown, this embodiment provides a method for matching a beautiful voice filter, including:

[0058] S1: Get the target audio;

[0059] S2: importing the target audio into a pre-built filter analysis model to obtain a target filter;

[0060] S3: Using a target filter, reverberate the target audio to generate a target reverberation audio.

[0061] The existing karaoke effects allow users to choose scenes. A scene is a combination of effect parameters set based on experience. There are also parameters open to users for adjustment. However, what can be switched are the effect processing process, not the final effect target. However, on the one hand, the preset parameters are manually adjusted based on the experience of the tuner. The style and aesthetics are closely related to people, and the direction tends to be fixed. On the other hand, users switch the effect processing process, not the final effect target. In order to achieve the final effect, different sounds require different processing.

[0062] Therefore, in this embodiment, through a large amount of learning of commercial music, a series of highly characteristic vocal presentation methods for different music styles are obtained and packaged into filters. When the user uses it, according to the actual situation of the user, the corresponding filter is automatically selected through the filter analysis model, and then applied to the target audio for reverberation. This allows the user to obtain a significant personalized voice beautification effect in real time when singing a song, so that the user can easily add the desired personalized beautification effect to his or her voice with one click.

[0063] At the same time, according to actual needs, when the user is actually singing, the user can select the corresponding filter and process the user's voice accordingly according to the user's characteristics.

[0064] In this embodiment, filters are selected and applied for the user's voice through automatic filter selection, which is independent of the tuner's experience and achieves a one-click beautification effect.

[0065] Embodiment 2:

[0066] like Figure 2 As shown, this embodiment provides a method for matching a beautiful voice filter, including:

[0067] S0-1: Use the sound source separation tool to extract the reverberant vocal data from the original sound data;

[0068] S0-2: performing reverberation separation on the reverberation vocal data to obtain pure vocal data and reverberation data;

[0069] S0-3: Classify and train the pure vocal data and the reverberation data to obtain multiple singer-vocal statistics and multiple song-reverberation statistics; the singer-vocal statistics include the frequency spectrum distribution, dynamic differential parameter distribution, and effect sound volume of the target singer; the song-reverberation statistics include the reverberation frequency spectrum distribution, reverberation time, and reverberation volume of the target song;

[0070] S0-4: Perform cluster analysis based on the multiple singer-vocal statistics and the multiple song-reverberation statistics to generate a filter analysis model.

[0071] S1: Get the target audio;

[0072] S2: importing the target audio into a pre-built filter analysis model to obtain a target filter;

[0073] S3: Using a target filter, reverberate the target audio to generate a target reverberation audio.

[0074] In the embodiment, the construction of the filter analysis model is generally achieved through three steps from S0-1 to S0-4. Specifically, firstly, a sound source separation tool is used to extract the reverberant vocal data in the original sound data. Generally speaking, the public database of the original sound data and the audio data in the commercial MP3 format purchased from the regular channels are used. For the audio data in the commercial MP3 format purchased from the regular channels, an open source deep learning sound source separation tool, such as sleepter, is generally used to extract the vocals from the original audio data.

[0075] Generally speaking, the database will consist of the pure vocal parts and pure accompaniment parts of a large number of songs of different styles and by different singers, so as to meet the data required for the process of building the filter analysis model.

[0076] In the specific model construction process, because the purpose of this embodiment is to form a filter, that is, the reverberation required in the singing process, the reverberation separation tool is first used to separate the reverberation in the sound from the human voice, and the reverberation is extracted to proceed to the next step. Then, the required filter analysis model is generated through the cluster analysis method to select which filter to use.

[0077] Further preferably, the reverberation separation specifically includes:

[0078] S0-2-1: Use linear prediction coding to perform reverberation separation to obtain pure vocal data and reverberation data;

[0079] like Figure 7 As shown, the cost function of the linear predictive coding is Wherein, R is an impulse response filter, O is the order of the impulse response filter, and N is the length of the impulse response filter.

[0080] Generally speaking, when using the LPC linear prediction method for reverberation separation, since the human voice has been separated as a whole track, there is no need to consider causality in this step. It is only necessary to calculate a finite impulse response filter that can minimize the signal energy.

[0081] For the separated pure human voice data v, we need to minimize the energy of Rv. That is, solve R so that the cost function is This is a standard linear prediction problem. The corresponding R can be solved using the standard LPC linear prediction method. The solved Rv is the de-coherent original sound, and v-Rv is the reverberated wet sound. The impulse response of the filter R is equivalent to the reverberation effect.

[0082] Further preferably, the length of the impulse response filter of the linear prediction coding is 0.5 seconds.

[0083] Considering that the reverb of commercial songs is generally not too long, we choose a half-second long filter R. For the separated vocal signal v, we require to minimize the energy of Rv.

[0084] Embodiment 3:

[0085] like Figure 3 As shown, this embodiment provides a method for matching a beautiful voice filter, including:

[0086] S0-1: Use the sound source separation tool to extract the reverberant vocal data from the original sound data;

[0087] S0-2-1: Use linear prediction coding to separate reverberation to obtain pure vocal data and reverberation data; the cost function of the linear prediction coding is Wherein, R is an impulse response filter, O is the order of the impulse response filter, and N is the length of the impulse response filter; the length of the impulse response filter of the linear prediction coding is 0.5 seconds.

[0088] S0-3-1: Divide the pure human voice data into pure human voice data sets according to a preset human voice block length;

[0089] S0-3-2: Divide the reverberation data into reverberation data sets according to a preset reverberation block length;

[0090] S0-3-3: Counting the pure vocal data set according to a preset vocal statistical algorithm to generate singer-vocal statistical results;

[0091] S0-3-4: Counting the reverberation data set according to a preset reverberation statistical algorithm to generate song-reverberation statistical results;

[0092] S0-4: performing cluster analysis based on the plurality of singer-voice statistics and the plurality of song-reverberation statistics to generate a filter analysis model;

[0093] S1: Get the target audio;

[0094] S2: importing the target audio into a pre-built filter analysis model to obtain a target filter;

[0095] S3: Using a target filter, reverberate the target audio to generate a target reverberation audio.

[0096] Because the original data has been classified by singer, we collect statistics for each singer. We assume that the voice of each singer is relatively consistent during singing. All the data of each singer is processed by partition blocks, and the reverberation data of each song is also processed by partition blocks, so as to generate the corresponding pure vocal data set and reverberation data set, and perform cluster analysis on the original pure vocal data set and reverberation data set, and then generate the corresponding filter analysis model.

[0097] Further preferably, the length of the vocal block is 0.5 seconds; the length of the reverberation block is 0.5 seconds, so as to adapt to the length of the impulse response filter.

[0098] Further preferably, the vocal statistics algorithm specifically includes: taking statistics on the MFCC coefficients and the average volume of the blocks and the average volume of the wet sound part of the target vocal statistics set to generate singer-vocal statistics results; the reverberation statistics algorithm specifically includes: taking statistics on the MFCC coefficients and the average volume of the blocks and the average volume of the wet sound part of the target reverberation statistics set to generate song-reverberation statistics results.

[0099] For each block in the pure human voice dataset, the following calculation is performed:

[0100] First, determine whether there is singing. If there is singing, then count the MFCC coefficients of the current block, and use the square average to represent the volume of the current block, and calculate the volume of the current block of the wet sound part.

[0101] Then, the MFCC and volume of all the singer's blocks are counted, and the spectrum distribution and dynamic range distribution of the singer, as well as the effect sound volume, are obtained. At the same time, for each song, since the reverberation response R is obtained, the spectrum distribution, reverberation time, and reverberation volume of the reverberation are also counted, thus obtaining the statistics of the reverberation.

[0102] Embodiment 4:

[0103] like Figure 4 As shown, this embodiment provides a method for matching a beautiful voice filter, including:

[0104] S0-1: Use the sound source separation tool to extract the reverberant vocal data from the original sound data;

[0105] S0-2: performing reverberation separation on the reverberation vocal data to obtain pure vocal data and reverberation data;

[0106] S0-3: Classify and train the pure vocal data and the reverberation data to obtain multiple singer-vocal statistics and multiple song-reverberation statistics; the singer-vocal statistics include the frequency spectrum distribution, dynamic differential parameter distribution, and effect sound volume of the target singer; the song-reverberation statistics include the reverberation frequency spectrum distribution, reverberation time, and reverberation volume of the target song;

[0107] S0-4-1: Importing the plurality of singer-voice statistics and the song-reverberation statistics into a K-means algorithm to generate a filter feature vector clustering set;

[0108] S0-4-2: selecting a preset number of clusters, from among the filter feature vector clusters, a filter feature vector cluster with the largest number of feature vectors, to generate a filter analysis model;

[0109] S1: Get the target audio;

[0110] S2: importing the target audio into a pre-built filter analysis model to obtain a target filter;

[0111] S3: Using a target filter, reverberate the target audio to generate a target reverberation audio.

[0112] After the process of S0-3, we get a timbre (spectral distribution), dynamic volume distribution, effect volume, reverberation timbre, and reverberation length for each singer, namely the singer-vocal statistics and song-reverberation statistics. Then for each singer, we form a vector that lists all the above features, which is the feature vector of the singer. Then we use the K-means algorithm to perform statistics, and finally get the distribution of the feature vector, which is the final filter analysis model. At the same time, we select several commonly used filters, that is, several clusters with a large number, for use in subsequent business processes.

[0113] Further preferably, the K value of the K-means algorithm is 2 times the preset number of clusters to be selected.

[0114] In the specific setting of K value, half of the clusters with a larger number are generally selected. For example, if N filters are needed in the end, the feature vectors of all singers are unsupervised clustered using the standard k-mean method to cluster into 2N clusters. The N clusters with the most elements are selected, and the average value of the features is used as the feature vector of the N filters.

[0115] Embodiment 5:

[0116] like Figure 5 As shown, this embodiment provides a method for matching a beautiful voice filter, including:

[0117] S1: Get the target audio;

[0118] S2-1: Divide the target audio into a target audio block set according to a preset target audio block length;

[0119] S2-2: Arrange the target audio block set in chronological order, and make the MFCC coefficient of the current audio block equal to the MFCC coefficient of the first target audio block in the target audio block set;

[0120] S2-3: counting the MFCC coefficients of the target audio block in sequence to update the MFCC coefficients of the current audio block; when updating the MFCC coefficients of the current audio block, a change in the MFCC coefficients of the current audio block is not greater than 1 decibel.

[0121] S3: Using a target filter, reverberate the target audio to generate a target reverberation audio.

[0122] In the specific use process, the built-in graphic EQ algorithm divides the target audio into several frequency bands, and each frequency band can independently control its gain, that is, different filters can be used in each different frequency band. The arrangement of the frequency bands is exactly the same as the frequency division bands of the MFCC coefficients, and the length is 0.5 seconds.

[0123] According to the user's singing, the user's MFCC coefficients are counted in real time. At the same time, in order to ensure the stability of the user's tone, the user's MFCC system statistics are slowly updated at a rate of up to 1db every 0.5 seconds. In this way, after the user sings for a period of time, the user's MFCC coefficients can be smoothly obtained, which is more reliable.

[0124] This embodiment achieves stable MFCC coefficient statistics by limiting the update changes of MFCC coefficients, reduces the interference of noise and fluctuations on filter selection, and improves the reliability and stability of the system.

[0125] At the same time, it is compared with the MFCC coefficients in the feature vector in real time, and the gain of each frequency band of the graphic EQ is adjusted according to the difference between the user MFCC and the target MFCC.

[0126] At the same time, since the reverberation impulse response R of the current style has been separated, we know how long the reverberation of the current style is needed, that is, the time for the impulse response to decay to -60db. The volume of the current reverberation, that is, the volume ratio of Rv to v, and the EQ of the current reverberation, that is, the spectrum envelope of R. Since this reverberation algorithm provides the adjustment of reverberation length, reverberation EQ, and reverberation volume, it only needs to be adjusted according to the feature vector.

[0127] Embodiment 6:

[0128] like Figure 6 As shown, this embodiment provides a method for matching a beautiful voice filter, including:

[0129] S0-1: Use the sound source separation tool to extract the reverberant vocal data from the original sound data;

[0130] S0-2-1: Use linear prediction coding to separate reverberation to obtain pure vocal data and reverberation data; the cost function of the linear prediction coding is Wherein, R is an impulse response filter, O is the order of the impulse response filter, and N is the length of the impulse response filter; the length of the impulse response filter of the linear prediction coding is 0.5 seconds.

[0131] S0-3-1: Divide the pure human voice data into pure human voice data sets according to a preset human voice block length;

[0132] S0-3-2: Divide the reverberation data into reverberation data sets according to a preset reverberation block length;

[0133] S0-3-3: Counting the pure vocal data set according to a preset vocal statistical algorithm to generate singer-vocal statistical results;

[0134] S0-3-4: Counting the reverberation data set according to a preset reverberation statistical algorithm to generate song-reverberation statistical results;

[0135] The singer-voice statistics include the target singer's spectrum distribution, dynamic difference parameter distribution, and effect sound volume; the song-reverberation statistics include the target song's reverberation spectrum distribution, reverberation time, and reverberation volume;

[0136] S0-4-1: Importing the plurality of singer-voice statistics and the song-reverberation statistics into a K-means algorithm to generate a filter feature vector clustering set;

[0137] S0-4-2: Select a preset number of clustering clusters, which is the filter feature vector clustering cluster with the largest number of feature vectors in the filter feature vector clustering cluster set, to generate a filter analysis model; the K value of the K-means algorithm is 2 times the preset number of clustering clusters.

[0138] S1: Get the target audio;

[0139] S2-1: Divide the target audio into a target audio block set according to a preset target audio block length;

[0140] S2-2: Arrange the target audio block set in chronological order, and make the MFCC coefficient of the current audio block equal to the MFCC coefficient of the first target audio block in the target audio block set;

[0141] S2-3: counting the MFCC coefficients of the target audio block in sequence to update the MFCC coefficients of the current audio block; when updating the MFCC coefficients of the current audio block, a change in the MFCC coefficients of the current audio block is not greater than 1 decibel.

[0142] S3: Using a target filter, reverberate the target audio to generate a target reverberation audio.

[0143] Through the above embodiments, the present invention achieves:

[0144] (1) Through automated filter selection, filters are selected and applied based on the user's voice, without relying on the tuner's experience, achieving a one-click beautification effect;

[0145] (2) By limiting the update and change of MFCC coefficients, stable MFCC coefficient statistics are achieved, reducing the interference of noise and fluctuations on filter selection, and improving the reliability and stability of the system.

[0146] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0147] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for matching a beautiful voice filter, characterized in that: include: Get the target audio; Importing the target audio into a pre-built filter analysis model to obtain a target filter, including: dividing the target audio into a target audio block set according to a preset target audio block length; Arrange the target audio block set in chronological order, and make the MFCC coefficient of the current audio block equal to the MFCC coefficient of the first target audio block in the target audio block set; sequentially counting the MFCC coefficients of the target audio block to update the MFCC coefficients of the current audio block; when updating the MFCC coefficients of the current audio block, a change in the MFCC coefficients of the current audio block is not greater than 1 decibel; The target audio is reverberated using a target filter to generate a target reverberated audio.

2. A method for matching a beautiful voice filter according to claim 1, characterized in that: Also includes: Constructing the filter analysis model specifically includes: Use sound source separation tools to extract reverberant vocal data from original sound data; Performing reverberation separation on the reverberation vocal data to obtain pure vocal data and reverberation data; Classification training is performed on the pure vocal data and the reverberation data to obtain a plurality of singer-vocal statistical results and a plurality of song-reverberation statistical results; the singer-vocal statistical results include the frequency spectrum distribution, dynamic differential parameter distribution, and effect sound volume of the target singer; the song-reverberation statistical results include the reverberation frequency spectrum distribution, reverberation time, and reverberation volume of the target song; A cluster analysis is performed based on the multiple singer-vocal statistics and the multiple song-reverberation statistics to generate a filter analysis model.

3. A method for matching beautiful voice filters according to claim 2, characterized in that: The reverberation separation specifically includes: Use linear predictive coding to perform reverberation separation to obtain pure vocal data and reverberation data; The cost function of the linear predictive coding is Wherein, R is an impulse response filter, O is the order of the impulse response filter, and N is the length of the impulse response filter.

4. A method for matching beautiful voice filters according to claim 3, characterized in that: The length of the impulse response filter of the linear prediction coding is 0.5 seconds.

5. A method for matching a beautiful voice filter according to any one of claims 2 to 4, characterized in that: The classification training specifically includes: Dividing the pure vocal data into pure vocal data sets according to a preset vocal block length; Dividing the reverberation data into reverberation data sets according to a preset reverberation block length; According to a preset vocal statistical algorithm, the pure vocal data set is counted to generate a singer-vocal statistical result; According to a preset reverberation statistics algorithm, the reverberation data set is counted to generate a song-reverberation statistics result.

6. A method for matching beautiful voice filters according to claim 5, characterized in that: The length of the vocal block is 0.5 seconds; The reverberation block length is 0.5 seconds.

7. A method for matching beautiful voice filters according to claim 5, characterized in that: The human voice statistics algorithm specifically includes: Get all vocal chunks of the target singer; Obtaining the vocal blocks in which the target singer is currently speaking from among all the vocal blocks, to form a target vocal statistics set; The MFCC coefficients, the average volume of each block, and the average volume of the wet sound part of the target vocal statistical set are counted to generate singer-vocal statistical results; The reverberation statistics algorithm specifically includes: Obtain all reverberation blocks of the target song to form a target reverberation statistics set; The MFCC coefficients, the average volume of each block, and the average volume of the wet sound part of the target reverberation statistical set are statistically analyzed to generate song-reverberation statistical results.

8. A method for matching beautiful voice filters according to claim 2, characterized in that: The cluster analysis specifically includes: Specifically, the method includes: importing the plurality of singer-voice statistics and the song-reverberation statistics into a K-means algorithm to generate a filter feature vector clustering set; A preset number of clusters are selected, and the filter feature vector cluster with the largest number of feature vectors in the filter feature vector cluster set is selected to generate a filter analysis model.

9. A method for matching beautiful voice filters according to claim 8, characterized in that: The K value of the K-means algorithm is 2 times the preset number of clusters to be selected.

Images