Graduate School and Research Center in Digital Sciences

ODESSA at Albayzin speaker diarization challenge 2018

Patino, Jose; Delgado, Héctor; Yin, Ruiqing; Bredin, Hervé; Barras; Claude; Evans, Nicholas

IBERSPEECH 2018, 21-23 November 2018, Barcelona, Spain

Best System Award IberSpeech 2018 Albayzin Challenges - Speaker Diarization Open Condition

This paper describes the ODESSA submissions to the Albayzin Speaker Diarization Challenge 2018. The challenge addresses the diarization of TV shows. This work explores three different techniques to represent speech segments, namely binary key, x-vector and triplet-loss based embeddings. While training-free methods such as the binary key technique can be applied easily to a scenario where training data is limited, the training of robust neural-embedding extractors is considerably more challenging. However, when training data is plentiful (open-set condition), neural embeddings provide more robust segmentations, giving speaker representations which lead to better diarization performance. The paper also reports our efforts to improve speaker diarization performance through system combination. For systems with a common temporal resolution, fusion is performed at segment level during clustering. When the systems under fusion produce segmentations with an arbitrary resolution, they are combined at diarization hypothesis level. Both approaches to fusion are shown to improve diarization performance.

Document Hal Bibtex

Title:ODESSA at Albayzin speaker diarization challenge 2018
Keywords:speaker diarization, diarization fusion, neural embeddings, binary key
Department:Digital Security
Eurecom ref:5732
Copyright: © ISCA. Personal use of this material is permitted. The definitive version of this paper was published in IBERSPEECH 2018, 21-23 November 2018, Barcelona, Spain and is available at :
Bibtex: @inproceedings{EURECOM+5732, year = {2018}, title = {{ODESSA} at {A}lbayzin speaker diarization challenge 2018}, author = {{P}atino, {J}ose and {D}elgado, {H}{\'e}ctor and {Y}in, {R}uiqing and {B}redin, {H}erv{\'e} and {B}arras and {C}laude and {E}vans, {N}icholas}, booktitle = {{IBERSPEECH} 2018, 21-23 {N}ovember 2018, {B}arcelona, {S}pain}, address = {{B}arcelona, {SPAIN}}, month = {11}, url = {} }
See also: