Amid digital music's growth, challenges like multi-track interference and noise robustness persist.Traditional single-domain analysis struggles with harmonic and transient details.We propose a computer music sound signal separation model (CMSSM-TFCFS), which includes an encoder with time-frequency cross-domain feature selection, a residual temporal convolution-based separator for long-term dependencies, and a decoder.It is jointly trained with an acoustic parameter synthesis model (APSM-NV) that uses a multilayer LSTM to predict clean acoustic features and a transformer-based vocoder for waveform generation.On a self-built dataset, the separation model achieves a signal-to-distortion ratio of 16.6 dB and a scale-invariant signal-to-noise ratio of 16.9 dB, improving baselines by 6.4% and 10.5%.By dynamically integrating time-frequency features and enabling end-to-end optimisation, this work offers a new paradigm for complex music signal processing, advancing support for music production and audio restoration, and promoting progress in digital music processing.
Yanyan Wang (Thu,) studied this question.