Overlapped voice detection method and system

Abstract

The invention provides an overlapped voice detection method and an overlapped voice detection system. The method comprises the following steps: based on the Bayes information criterion, finding voice fractions which only contain voice of a single speaking person from a plurality of voice fractions in the overlapped voice, and assigning an identical identifier to voice fractions belonging to the same speaking person; randomly selecting the sample data of various voice fractions from the same type of voice fractions and combining the selected sample data so as to obtain various combination results capable of reflecting all voice overlapping possibilities; establishing a single-person speech fraction model and an overlapped voice fraction model on the basis of the obtained single-person voice fractions and the combined multi-person overlapped voice fractions; and finally detecting each voice fraction by using the single-person voice fraction model and the overlapped voice fraction model, and labeling each voice fraction according to the detection result.

Description

technical field [0001] The invention relates to a method and a system for marking a speaker in a segment of speech, in particular to identifying a speaker in a segment of speech. Background technique [0002] In some practical applications, it is usually necessary to recognize a piece of speech to know how many people are speaking in the speech and which part of the speech is said by that person. This recognition is especially important when there are multiple speakers in a piece of speech. At this stage, especially in conferences, overlapping speech (there are multiple people speaking) is a very important source of error when it comes to labeling speakers. Current speech annotation systems struggle to correctly identify overlapping speech segments that contain multiple speakers. Existing speech segment recognition systems usually can only recognize speech segments containing a single speaker for each segment of speech, and for speech segments containing multiple speakers,...

AI Technical Summary

Problems solved by technology

US7295970B discloses a method of training and using a separate speaker model to train overlapping voices, although the patent also mentions a method for discovering overlapping voices and can separate overlapping voices from separate voices, but the patent disclosed The method cannot identify who is speaking in an overlapping speech

[0004] US7646894B2 discloses a method of using a Bayesian competition model to identify an unmarked speaker appearing in a section of speech, however, this patented method cannot handle overlapping speech

[0005] Therefore, the problem in the prior art is that it is impossible to obtain information about who is speaking and when speaking for any segment of speech.

Images