Skip to main navigation Skip to search Skip to main content

Randomization effect on iterative-based speaker diarization system for telephone conversations

  • Tal Furmanov
  • , Lidiya Aminov
  • , Ami Moyal
  • , Itshak Lapidot

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

The primary objective of speaker diarization system is to designate speech segments to one of K speakers in the conversation. We use a hidden-distortion-model (HDM)-based system. HDM allows using different emission models as speaker models. We investigate the effect of randomization in two different levels. One level is stochastic training versus deterministic training and the other, random model initialization versus preserving initialization from the previous iteration. The emission models were codebooks (CBs) trained using K-means algorithm, both, batch and stochastic versions, as well as a self-organizing map (SOM) in its stochastic version. The evaluation performed on 108 telephone conversations from the LDC CallHome corpus. We will show that randomizing is always outperforming the deterministic training. Stochastic training demonstrated relative improvement of 3.5%. Random initialization achieved relative improvement of 7.28% comparing to preservation of initialization from the previous iteration.

Original languageEnglish
Title of host publication2014 IEEE 28th Convention of Electrical and Electronics Engineers in Israel, IEEEI 2014
PublisherInstitute of Electrical and Electronics Engineers
ISBN (Electronic)9781479959877
DOIs
StatePublished - 1 Jan 2014
Externally publishedYes
Event2014 28th IEEE Convention of Electrical and Electronics Engineers in Israel, IEEEI 2014 - Eilat, Israel
Duration: 3 Dec 20145 Dec 2014

Publication series

Name2014 IEEE 28th Convention of Electrical and Electronics Engineers in Israel, IEEEI 2014

Conference

Conference2014 28th IEEE Convention of Electrical and Electronics Engineers in Israel, IEEEI 2014
Country/TerritoryIsrael
CityEilat
Period3/12/145/12/14

Keywords

  • Hidden-distortion model (HDM)
  • Initialization
  • K-means
  • Self-organizing maps (SOM)
  • Speaker diarization

ASJC Scopus subject areas

  • Electrical and Electronic Engineering

Fingerprint

Dive into the research topics of 'Randomization effect on iterative-based speaker diarization system for telephone conversations'. Together they form a unique fingerprint.

Cite this