Multi-speaker-voice separation method based on deep learning

Abstract

The invention discloses a multi-speaker-voice separation method based on deep learning. The method comprises the following steps: performing framing, windowing and Fourier transform on multi-speaker mixed voice signals obtained by a microphone to obtain a frequency spectrum of the mixed signals; sending frequency spectrum amplitude of the mixed voice signals to a neural network and estimating ideal amplitude masking of each target signal; retrieving the phase of each target signal with an iteration method on the basis of the estimated ideal amplitude masking of the signal as well as the frequency spectrum amplitude and frequency spectrum phase of the mixed signals; calculating phase-sensitive masking of each target signal by the retrieved phase of the target signal, and training the neuralnetwork to estimate the phase-sensitive masking; obtaining frequency spectrum amplitude of each target signal according to the phase-sensitive masking estimated by the neural network, and reconstructing the frequency spectrum of each target signal through combination with the retrieved phase; performing Fourier transform on the reconstructed frequency spectra of the signals to obtain separated time-domain voice signals. The speaker voice separation effect can be improved effectively with the method.

Description

technical field [0001] The present invention relates to the technical field of voice separation, in particular to a deep learning-based multi-person voice separation method. Background technique [0002] Speaker separation technology is to extract the speech signal of each speaker from the mixed speech signals of multiple speakers. This technology is of great significance to target speaker detection, speech recognition, etc. [0003] Due to the complexity and instability of the speech signal, the traditional separation method can not achieve a good separation effect, and the previous separation only estimates the spectrum amplitude of the target signal. Contents of the invention [0004] The purpose of the present invention is to solve the defects in the prior art. [0005] In order to achieve the above object, a deep learning-based multi-person voice separation method includes steps: [0006] Framing, windowing, and Fourier transforming the mixed voice signal containin...

AI Technical Summary

Problems solved by technology

[0003] Due to the complexity and instability of the speech signal, the traditional separation method cannot achieve a good separation effect, and in the previous separation, only the spectral amplitude of the target signal was estimated

Images