Audio data processing

1 . A method of media processing, comprising:

extracting, by processing circuitry of a server device, audio track data of at least a signal source type from audio data, the audio data comprising multiple data segments, the audio track data comprising at least a time period that is determined to be related to the signal source type;

allocating weight values respectively to the data segments in the audio data according to the audio track data, wherein when the audio track data comprises a first signal source type of voice, a first weight value for a first data segment in a first time period related to the voice is determined based on a voice parameter of the first data segment, the first weight value is positively correlated with the voice parameter, and the voice parameter comprises at least one of speech speed, intonation, and/or loudness;

concatenating the weight values to form a weight value sequence of the audio data;

extracting audio features respectively from the data segments;

concatenating the audio features of the data segments to form an audio feature sequence of the audio data;

encoding the audio feature sequence to obtain an attention parameter sequence of the audio data;

fusing the attention parameter sequence and the weight value sequence to obtain respective fusion parameters for the data segments;

determining recommendation parameters of the data segments based on the fusion parameters;

determining, by the processing circuitry of the server device, a recommended segment in the audio data based on the recommendation parameters of the data segments; and

transmitting, from the server device to a terminal device that plays media including the audio data, signals indicative of the recommended segment in the audio data.

2 . The method according to claim 1 , wherein the extracting the audio track data comprises:

extracting a global feature of the audio data from the audio data; and

performing a signal source separation on the audio data by using the global feature as a mask to obtain the audio track data, a boundary in the mask representing a boundary between different signal source types.

3 . The method according to claim 2 , wherein the extracting the global feature of the audio data from the audio data comprises:

performing a feature extraction on the audio data to obtain an original feature of the audio data;

performing multiple levels of pooling processing on the original feature to obtain multiple local features of the audio data; and

concatenating the multiple local features to obtain the global feature of the audio data.

4 . The method according to claim 1 , wherein the extracting the audio track data further comprises:

determining that the first time period is related to the voice when short time energy in the first time period is greater than an energy threshold and a zero cross counter in the first time period is less than a zero cross counter threshold.

5 . The method according to claim 1 , wherein the extracting the audio track data further comprises:

determining, when the audio track data comprises a second signal source type of a background sound, that a second time period is related to the background sound when the audio data in the second time period satisfies at least one of:

loudness of the second time period is greater than a lower loudness limit; and/or

a length of the second time period is greater than a lower length limit.

6 . The method according to claim 1 , further comprising at least one of:

allocating a preset value as a second weight value to a second data segment in a second time period related to a second signal source type of background sound, the preset value being less than the first weight value of the first data segment related to the voice; and/or

allocating zero as a third weight value to a third data segment when the third data segment does not belong to a time period related to the first signal source type or related to the second signal source type.

7 . The method according to claim 1 , wherein the audio track data comprises a single signal source type of background sound, and the allocating the weight values to the data segments comprises:

determining, based on a background sound parameter of a data segment in a first time period related to the background sound, a weight value corresponding to the data segment, the weight value being positively correlated with the background sound parameter, and the background sound parameter comprising at least one of loudness and/or tone; or

allocating zero as the weight value to the data segment when the data segment does not belong to a time period related to the single signal source type of the background sound.

8 . The method according to claim 1 , wherein further comprising:

for each data segment in the data segments,

extracting a time-domain signal feature and a frequency-domain signal feature of the respective data segment;

determining, based on a first time period related to the signal source type, a one-dimensional time-domain weight value corresponding to the time-domain signal feature; and

determining, based on the first time period related to the signal source type, a two-dimensional frequency-domain weight value corresponding to the frequency-domain signal feature;

performing multiple levels of convolution on a product of one-dimensional time-domain weight values of the data segments and time-domain signal features of the data segments to obtain time-domain audio features;

performing multiple levels of convolution on a product of two-dimensional frequency-domain weight values of the data segments and frequency-domain signal features of the data segments to obtain frequency-domain audio features;

resizing the time-domain audio features to obtain two-dimensional time-domain audio features; and

fusing the two-dimensional time-domain audio features with the frequency-domain audio features to obtain the audio features of the data segments.

9 . The method according to claim 8 , wherein the fusing the two-dimensional time-domain audio features with the frequency-domain audio features to obtain the audio features of the data segments comprises:

determining a superposition feature of a two-dimensional time-domain audio feature and a frequency-domain audio feature for a data segment of the data segments;

performing two-dimensional convolution on the superposition feature to obtain a two-dimensional convolution result;

obtaining a maximum superposition feature and a mean superposition feature of the two-dimensional convolution result; and

performing linear activation on a sum of the maximum superposition feature and the mean superposition feature to obtain an audio feature of the data segment.

10 . The method according to claim 1 , wherein the encoding the audio feature sequence to obtain the attention parameter sequence of the audio data comprises:

for a first audio feature in the audio feature sequence that is an audio feature of a first data segment in the data segments:

respectively fusing the first audio feature and each audio feature of other data segments in the data segments based on an attention mechanism to obtain weighted correlation degrees corresponding to the first audio feature; and

adding the weighted correlation degrees to obtain an attention parameter corresponding to the first audio feature; and

concatenating attention parameters corresponding to the audio features based on a sequence of the data segments to form the attention parameter sequence of the audio data.

11 . The method according to claim 10 , further comprising:

performing full connection on an audio feature of each data segment of the audio data to obtain an embedding vector of each audio feature.

12 . The method according to claim 10 , wherein:

the respectively fusing the first audio feature and each audio feature of the other data segments comprises:

multiplying a first embedding vector of the first audio feature of the first data segment by a second embedding vector of a second audio feature of another data segment to obtain a correlation degree between the first audio feature and the second audio feature of the other data segment; and

multiplying the first audio feature by the correlation degree to obtain a weighted correlation degree corresponding to the first audio feature.

13 . The method according to claim 1 , wherein the fusing the attention parameter sequence and the weight value sequence to obtain the fusion parameters respectively for the data segments further comprises:

obtaining, from the attention parameter sequence, an attention parameter corresponding to a data segment; and

multiplying a weight value of the data segment by the attention parameter of the data segment to obtain a fusion parameter of the data segment; and

the determining the recommendation parameters of the data segments based on the fusion parameters comprises:

performing normalization processing on a fusion parameter of a data segment to obtain a recommendation parameter of the data segment.

14 . The method according to claim 1 , wherein the determining the recommended segment in the audio data based on the recommendation parameters of the data segment comprises:

sorting the data segments in a descending order based on the recommendation parameters of the data segments; and

determining at least one data segment from a head portion of the sorted data segments as the recommended segment of the audio data.

15 . The method according to claim 1 , wherein the determining the recommended segment in the audio data based on the recommendation parameters of the data segments comprises:

determining a data segment whose recommendation parameter is greater than a recommendation parameter threshold as the recommended segment.

16 . The method according to claim 1 , further comprising:

generating a recommendation parameter curve of the audio data based on the recommendation parameters of the data segments; and

displaying the recommendation parameter curve of the audio data on a playback interface in response to a playback trigger operation, the recommendation parameter curve representing a relationship of the recommendation parameters to a playback time of the audio data.

17 . The method according to claim 16 , further comprising:

displaying a label of the recommended segment on the playback interface, the label representing a time period of the recommended segment; and

jumping, in response to a selection operation for the label, to a start point of the recommended segment corresponding to the label to start playback.

18 . The method according to claim 1 , further comprising:

training an audio processing model based on sample audio data to obtain a trained audio processing model; and

performing the media processing based on the trained audio processing model, wherein the training the audio processing model comprises:

concatenating, based on label values of actual recommended segments of the sample audio data, to form an actual recommendation parameter sequence of the sample audio data;

concatenating, based on predicted recommendation parameters of data segments of the sample audio data according to the audio processing model, to form a predicted recommendation parameter sequence of the sample audio data;

obtaining a cross-entropy loss of the audio processing model based on the actual recommendation parameter sequence and the predicted recommendation parameter sequence;

dividing the cross-entropy loss by a quantity of the data segments of the sample audio data to obtain a mean prediction loss; and

performing back propagation processing on the audio processing model based on the mean prediction loss to obtain an updated audio processing model.

19 . An apparatus for media processing, comprising processing circuitry configured to:

extract audio track data of at least a signal source type from audio data, the audio data comprising multiple data segments, the audio track data comprising at least a time period that is determined to be related to the signal source type;

allocate weight values respectively to the data segments in the audio data according to the audio track data, wherein when the audio track data comprises a first signal source type of voice, a first weight value for a first data segment in a first time period related to the voice is determined based on a voice parameter of the first data segment, the first weight value is positively correlated with the voice parameter, and the voice parameter comprises at least one of speech speed, intonation, and/or loudness;

concatenate the weight values to form a weight value sequence of the audio data;

extract audio features respectively from the data segments;

concatenate the audio features of the data segments to form an audio feature sequence of the audio data;

encode the audio feature sequence to obtain an attention parameter sequence of the audio data;

fuse the attention parameter sequence and the weight value sequence to obtain respective fusion parameters for the data segments;

determine recommendation parameters of the data segments based on the fusion parameters; and

determine a recommended segment in the audio data based on the recommendation parameters of the data segments.