Technical Improvements of the E-HMM Based Speaker Diarization System for Meeting Records

Abstract

International audienceThis paper is concerned with the speaker diarization task in the specific context of the meeting room recordings. Firstly, different technical improvements of an E-HMM based system are proposed and evaluated in the framework of the NIST RT'06S evaluation campaign. Related experiments show an absolute gain of 6.4% overall speaker di-arization error rate (DER) and 12.9% on the development and evaluation corpora respectively. Secondly, this paper presents an original strategy to deal with the overlapping speech. Indeed, speech overlaps between speakers are largely involved in meetings due to the spontaneous nature of this kind of data and they are responsible for a decrease in performance of the speaker di-arization system, if they are not dealt with. Experiments still conducted in the framework of the NIST RT'06S evaluation show the ability of the strategy in detecting overlapping speech (decrease of the missed speaker error rate), even if an overall gain in speaker diarization performance has not been achieved yet

Speaker Diarization: A Review of Recent Research

“…Model-based approaches tend to have better performances and rely on a two-class detector, with models pre-trained with external speech and non-speech data [6], [41], [49], [51], [52]. Speech and non-speech models may optionally be adapted to specific meeting conditions [15].…”

Identification of Soundbite and Its Speaker Name Using Transcripts of Broadcast News Speech

“…Gauvain et al [1998] and Sinha et al [2005] represented each fixed length window with a Gaussian distribution and measured the distance between those windows using divergence. The clustering step is achieved by a bottom-up fashion, that is, agglomerative clustering [Gauvain et al 1998;Ajmera and Wooters 2003;Reynolds and Torres-Carrasquillo 2004;Ben et al 2004], or a top-down manner, that is, divisive clustering (e.g., evolutive hidden Markov models in Fredouille and Senay [2006]). …”

Quality of service modeling and analysis for carrier ethernet

“…Mostly metric-based segmentation and clustering is being used [Cas04,vL06, [Cas04] have been used for segmentation or clustering. A hierarchical top-down clustering approach was taken by [FS07] using an HMM adding new states (representing speakers) at each iteration while re-segmenting the speech data. In [ZBLG07, AWP07] a bottom-up approach was taken, also using an HMM to realign the data after each iteration.…”

“…Some systems just picked a single channel [vL06,ZBLG07], while others segmented each channel separately [JLSW04] before combining the results or performed some form of pre-processing [FS07,AWP07] to combine the channels into one single recording.…”