Multistream speaker diarization beyond two acoustic feature streams

Abstract

Speaker diarization for meetings data are recently converging towards multistream systems. The most common complementary features used in combination with MFCC are Time Delay of Arrival (TDOA). Also other features have been proposed although, there are no reported improvements on top of MFCC+TDOA systems. In this work we investigate the combination of other feature sets along with MFCC+TDOA. We discuss issues and problems related to the weighting of four different streams proposing a solution based on a smoothed version of the speaker error. Experiments are presented on NIST RT06 meeting diarization evaluation. Results reveal that the combination of four acoustic feature streams results in a 30% relative improvement with respect to the MFCC+TDOA feature combination. To the authors' best knowledge, this is the first successful attempt to improve the MFCC+TDOA baseline including other feature streams.

Multistream speaker diarization through Information Bottleneck system outputs combination

“…Towards this end, the paper exthends the Information Bottleneck combination [6]. The method forms the space of relevance variables, necessary for the aIB clustering, using the output of two separate diarization systems.…”

“…6 12.4 aIB model-based combination outperforms the HMM/GMM system by 0.8% achieving state-of-the-art results; this system will be used as baseline and the proposed techniques will be benchmarked w.r.t. it.…”

“…we introduced a multi-stream diarization method based on the Information Bottleneck (IB) principle [6]. The combination of multiple streams happens using intermediate variables that carry relevant information about the problem, referred as relevance variables.…”

“…The combination of multiple streams happens using intermediate variables that carry relevant information about the problem, referred as relevance variables. The system was shown to outperform conventional HMM/GMM systems and being effective in combining up to four different feature streams [6]. This paper proposes a novel combination framework based on the IB principle which aims at combining the output of multiple diarization systems.…”

“…This paper proposes a novel combination framework based on the IB principle which aims at combining the output of multiple diarization systems. This novel combination will be referred as system based combination in contrast with what previously proposed in [6] referred as model based combination. Furthermore the paper introduces an hybrid model-system combination similar to the piped approaches described in [3] and [5].…”