Voice separation method, voice separation device, electronic equipment and storage medium

Abstract

The application provides a speech separation method, a speech separation device, electronic equipment, and a storage medium. The speech separation method includes: obtaining the original audio, and extracting the spectrogram feature sequence from the original audio in a time window sliding window; The graph feature sequence is input into the pre-trained speech segmentation model, and the embedded feature sequence is obtained through the speech segmentation model; the embedded feature sequence is input into the pre-trained speech clustering model, and the corresponding embedding feature sequence is obtained through the speech clustering model The predicted label sequence; perform single-speaker voice restoration based on the predicted label sequence to generate separated speech. According to the speech separation method, speech separation device, electronic equipment and storage medium of the present application, the problem of unsatisfactory speech separation effect can be solved, and the speech segment belonging to a single speaker can be separated from the short-term speech audio file in which multiple people speak alternately, And it can accurately estimate the number of speakers in conjunction with contextual information.

Description

technical field [0001] The present application relates to the technical field of voice processing, in particular to a voice separation method, a voice separation device, electronic equipment and a storage medium. Background technique [0002] "Cocktail party problem" is a classic problem in the field of computer speech processing. "Cocktail party problem" refers to that when a single speaker speaks, speech recognition technology can often accurately identify what the speaker is saying, but when the scene contains multiple When there are no speakers, the accuracy of speech recognition will be greatly reduced. [0003] Generally speaking, the audio to be separated will contain a large number of short-duration speech. Since the short-duration speech contains less information and is not highly distinguishable, it is difficult to separate the speech, and because the context information cannot be used to accurately estimate the number of speakers, the separation is difficult. The...

AI Technical Summary

Problems solved by technology

[0004] In view of the fact that the existing speech recognition technology is difficult to separate short-term speech and cannot accurately estimate the number of speakers in connection with context information, the application provides a speech separation method, speech separation device, electronic equipment and storage media

Images