Song training data processing method and device and computer readable storage medium

Abstract

The invention relates to the technical field of voice semantics, and discloses a song training data processing method. The song training data processing method comprises the steps that initial sampledata are acquired, the initial sample data comprise the music score of each song and the singing record corresponding to the music score of each song, the initial sample data are marked, training dataare obtained, and an acoustic feature output model is obtained through training based on the training data. Afterwards, the method processes the target training data, takes the linguistic feature andthe music feature of each song in the target training data as input data of the trained acoustic feature output model, outputs the acoustic feature of each song, and performs pitch transfer on each song according to the acoustic feature of each song and the music feature of each song. The invention further provides a song training data processing device and a computer readable storage medium. According to the method and the device, the number of training samples is increased under the condition that extra corpora is not recorded.

Description

technical field [0001] The invention relates to the technical field of speech semantics, in particular to a song training data processing method, device and computer-readable storage medium. Background technique [0002] The concept of singing synthesis has attracted people's attention since it was introduced. Its ultimate goal is to allow machines to sing songs of various melodies with a natural degree comparable to that of a real singer. Parametric synthesis is one of the mainstream technologies of singing synthesis. Its technical core is to let the model learn how to convert the language features of lyrics and music features of music scores into the acoustic features of singing through training models. Therefore, model training is a crucial step in parameter synthesis technology, and the performance of the trained model depends on the quality of the training corpus. If some contextual factors rarely or never appear in the training corpus, the model will not be able to le...

AI Technical Summary

Problems solved by technology

In addition, among all musical factors, pitch is one of the main reasons affecting singing quality, so making the corpus evenly and comprehensively summarizing each pitch is a major point when training data. In the prior art, many training samples need to be recorded. In order to get a comprehensive summary of the corpus of each pitch, it not only requires a lot of manpower, material resources and time, but also the pitch distribution is uneven

Images