This work proposes a dominance detection framework operating in reverberated environments. The framework is composed of a speech enhancement front-end, which automatically reduces the distortions introduced by room reverberation in the speech signals, and a dominance detector, which processes the enhanced signals and estimates the most and least dominant person in a segment. The front-end is composed by three cooperating blocks: speaker diarization, room impulse responses identification and speech dereverberation. The dominance estimation algorithm is based on bidirectional Long Short-Term Memory networks which allow for context-sensitive activity classification from audio feature functionals extracted via the real-time speech feature extraction toolkit openSMILE. Experiments have been performed suitably reverberating the DOME dataset: the absolute accuracy improvement averaged over the addressed reverberated conditions is 32.68% in the most dominant person estimation task and 36.56% in the least dominant person estimation one, both with full agreement among annotators.
Dominance Detection in A Reverberated Acoustic Scenario / Principi, Emanuele; R., Rotili; M., Woellmer; Squartini, Stefano; B., Schuller. - Volume 7367 LNCS, Issue PART 1:(2012), pp. 394-402. [10.1007/978-3-642-31346-2_45]
Dominance Detection in A Reverberated Acoustic Scenario
PRINCIPI, EMANUELE;SQUARTINI, Stefano;
2012-01-01
Abstract
This work proposes a dominance detection framework operating in reverberated environments. The framework is composed of a speech enhancement front-end, which automatically reduces the distortions introduced by room reverberation in the speech signals, and a dominance detector, which processes the enhanced signals and estimates the most and least dominant person in a segment. The front-end is composed by three cooperating blocks: speaker diarization, room impulse responses identification and speech dereverberation. The dominance estimation algorithm is based on bidirectional Long Short-Term Memory networks which allow for context-sensitive activity classification from audio feature functionals extracted via the real-time speech feature extraction toolkit openSMILE. Experiments have been performed suitably reverberating the DOME dataset: the absolute accuracy improvement averaged over the addressed reverberated conditions is 32.68% in the most dominant person estimation task and 36.56% in the least dominant person estimation one, both with full agreement among annotators.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.