專利授權區國立清華大學國際產學營運總中心 Operations Center for Industry Collaboration

搜尋專利授權區

關鍵字

» 新增關鍵字

選單

專利授權區

專利授權區
專利名稱(中)	唱歌技巧音訊轉換方法及其系統
專利名稱(英)	SINGING TECHNIQUE AUDIO CONVERTING METHOD AND SYSTEM THEREOF
專利家族	中華民國：I854481
專利權人	國立清華大學 100%
發明人	蘇豐文,陳柏維
技術領域	文化創意,資訊工程

專利摘要(英)
A singing technique audio converting method is proposed, and includes performing a spectrum converting step, a feature extracting step, a feature converting step, a classification comparing step, an optimal model establishing step and a singing technique converting step. The spectrum converting step includes configuring a processing module to convert an anchor spectrum data to a positive spectrum data according to a generative adversarial network model, and to generate a first loss function. The feature extracting step includes configuring the processing module to process the anchor spectrum data, the positive spectrum data and a negative spectrum data according to a self-supervised learning network model to extract a plurality of technique features and a plurality of content features. The feature converting step includes configuring the processing module to convert the content features to a plurality of output features according to a first branch model of the self-supervised learning network model, and to perform a loss operation procedure on the output features to generate a second loss function. The classification comparing step includes configuring the processing module to generate a plurality of classification probabilities by comparing two of the technique features according to the second branch model of the self-supervised learning network model, and to perform a cross-entropy operation procedure on the classification probabilities to generate a third loss function. The optimal model establishing step includes configuring the processing module to add the first loss function, the second loss function and the third loss function to generate a full loss function, and to adjust the generative adversarial network model according to the full loss function to establish an optimal generative adversarial network model. The singing technique converting step includes configuring the processing module to convert an audio data to a technique audio data by using the optimal generative adversarial network model. Therefore, the singing technique audio converting method of the present disclosure can convert a normal audio to an audio with a singing technique effectively and retain the content information of the normal audio completely.

專利摘要(英)

A singing technique audio converting method is proposed, and includes performing a spectrum converting step, a feature extracting step, a feature converting step, a classification comparing step, an optimal model establishing step and a singing technique converting step. The spectrum converting step includes configuring a processing module to convert an anchor spectrum data to a positive spectrum data according to a generative adversarial network model, and to generate a first loss function. The feature extracting step includes configuring the processing module to process the anchor spectrum data, the positive spectrum data and a negative spectrum data according to a self-supervised learning network model to extract a plurality of technique features and a plurality of content features. The feature converting step includes configuring the processing module to convert the content features to a plurality of output features according to a first branch model of the self-supervised learning network model, and to perform a loss operation procedure on the output features to generate a second loss function. The classification comparing step includes configuring the processing module to generate a plurality of classification probabilities by comparing two of the technique features according to the second branch model of the self-supervised learning network model, and to perform a cross-entropy operation procedure on the classification probabilities to generate a third loss function. The optimal model establishing step includes configuring the processing module to add the first loss function, the second loss function and the third loss function to generate a full loss function, and to adjust the generative adversarial network model according to the full loss function to establish an optimal generative adversarial network model. The singing technique converting step includes configuring the processing module to convert an audio data to a technique audio data by using the optimal generative adversarial network model. Therefore, the singing technique audio converting method of the present disclosure can convert a normal audio to an audio with a singing technique effectively and retain the content information of the normal audio completely.

聯絡資訊
承辦人姓名	李曉琪
承辦人電話	03-5715131 #31061
承辦人Email	hsiaochi@mx.nthu.edu.tw