Extracting single talker segments from audio mixtures in reverberant environments

Classifying the number of simultaneous speakers in a reverberant environment remains a difficult problem to solve. The less complicated but no less important problem of detecting whether a single speaker or multiple speakers are present also poses a challenge, as multipath often distorts or alters the features commonly used in speaker number classification. When this classification system is used as a front-end to a learning-based system, then it becomes imperative that the classifier be accurate. This thesis presents SinguDetect, an end-to-end system which uses the inter-aural phase differences between a pair of in-ear microphones to classify each 64 ms time window of audio as belonging to one speaker or not. Even in heavily reverberant environments (RT60 = 2.0 s), SinguDetect is still able to maintain a precision of around 0.8 where the number of simultaneous sources is not higher than three, whereas other state-of-the-art algorithms tend to suffer decreases in precision and f-score in this range.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Subramaniam, Avinash

Contributors dc:contributor

• Choudhury, Romit Roy

• Binaural
• Reverberant
• Classification

Statement dc:rights

• Copyright 2023 Avinash Subramaniam

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121540

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121540