The invention provides a training method, a synthesis method and a device for a high-expressive singing sound synthesis model. The method comprises the following steps: acquiring a lyric text and a music
score sequence obtained by predicting real singing sound; converting the lyric text into phoneme information, extracting
physical information of the singing sound from the singing sound audio data, and analyzing the
physical information to obtain intensity information; the phoneme information and attribute information of the music
score sequence are input into a text
encoder to generate a text embedding sequence, the duration of each word is determined by a duration predictor, and the text embedding sequence and the duration are input into a length constraint device to generate voice representation; inputting the speech representation into a
fundamental frequency decoder and a Mel decoder to generate a
fundamental frequency and a Mel spectrum; and adding the
fundamental frequency, the Mel spectrum and the voice representation to obtain a mixed embedded vector, inputting the mixed embedded vector into a
Transformer and a continuous voice unit representation module to generate a continuous voice unit vector, and synthesizing a predicted singing sound by a vocoder. According to the method provided by the invention, fine-grained modeling is carried out on each attribute,
coupling is avoided, and high-quality singing sound is synthesized.