Training method of song representation output model and song processing method

CN119479625AActive Publication Date: 2025-02-18TENCENT MUSIC ENTERTAINMENT TECH (SHENZHEN) CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411439840.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-15
Publication Date
2025-02-18
Estimated Expiration
2044-10-15

AI Technical Summary

Technical Problem

During the training of cross-modal song representation big model based on contrast learning methods, there may be semantic correlation between the text information of the song in the randomly selected negative sample and the text information of the song in the positive sample, resulting in poor model training results .

Method used

By obtaining song training data for multiple sample pairs, each sample pair includes song audio and song description text of the same song, for the current sample pair, according to the text difference between the song description text and the song description text in other sample pairs, Select song description text with relatively large text differences from other sample pairs as negative samples, and conduct comparative learning training on the song characterization output model based on the representation distance between these negative samples and positive samples.

Benefits of technology

By optimizing the selection of negative samples, the semantic correlation between negative samples and positive samples is avoided, the stability and effect of model training is improved, and the performance of song characterization output models is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119479625A_ABST
    Figure CN119479625A_ABST
Patent Text Reader

Abstract

The invention relates to a training method of a song representation output model and a song processing method. The method comprises the following steps: acquiring song training data of a plurality of sample pairs; for the current sample pair, according to the text difference between the song description text in the current sample pair and the song description text in each other sample pair, selecting a plurality of song description texts with relatively large text difference from each other sample pair as negative samples of the song audio in the current sample pair; determining a first characterization distance between the song audio in the current sample pair and the song description text, and determining a second characterization distance between the song audio in the current sample pair and a plurality of negative samples; and based on the first characterization distance and the second characterization distance, performing comparative learning training on the song characterization output model until a trained song characterization output model is obtained. By adopting the method, the problem of semantic similarity between the negative sample and the positive sample can be solved, and the model training effect and the model performance are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Acoustic event detection method and device, electronic equipment and storage medium

    CN114898737A

  • Speech recognition model training method, speech recognition method, equipment and medium

    CN117594037A

  • Model training method and device, electronic equipment and storage medium

    CN118098274A

  • Song recommendation model training method, computer equipment and storage medium

    CN118585668A

  • Semantic feature extraction model training method and device, equipment and storage medium

    CN118657151A