Single-channel voice separation method and device for multiple speakers

• Main Authors: SHI HUIYU, OUYANG PENG, YIN SHOUYI
• Format: Patent
• Language: Chinese; English
• Published: 05.02.2021
• Subjects: ACOUSTICS; MUSICAL INSTRUMENTS; PHYSICS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

The invention relates to a single-channel voice separation method and device for multiple speakers. The method comprises the following steps: acquiring a frequency spectrum amplitude and a frequency spectrum phase of mixed voice; and inputting the frequency spectrum amplitude of the mixed voice into a trained generative adversarial network model, and obtaining a plurality of estimated amplitude masks of the generative adversarial network model; obtaining a plurality of target frequency spectrum amplitudes according to the plurality of estimated amplitude masks and the frequency spectrum amplitude of the mixed voice; and reconstructing a plurality of target frequency spectrum amplitudes and frequency spectrum phases one by one to generate a plurality of target voices. According to the method, the target voice separation result corresponding to each speaker can be obtained, the number of speakers in the mixed voice can be quickly judged, the separation accuracy is improved, the distortion rate of the voice is r