Blind spatial sound source clustering and activity detection using uncalibrated microphone array

Abstract

Abstract-This paper presents a method for estimating the number, as well as the activity periods of spatially distributed sound sources using an uncalibrated microphone array. This methodology is applied for the purposes of speaker diarization. In general, speaker diarization has difficulty with: 1) estimating the number of sound sources (speakers), and 2) activity detection of multiple sound sources including overlap of utterances. Several microphone array based techniques have already tackled these challenges. However, existing methods mainly assume that the steering vectors for the microphone array are calibrated in advance to identify sound sources, which is difficult to satisfy when ad-hoc or flexible microphone arrays are used. Thus our approach estimates the number of sound sources blindly in two steps. First, Time Delay of Arrival (TDOA) of the observed signal is clustered. Second, the sound source activity is detected by clustering the long-term spatial spectrum using the TDOA based steering vector for each cluster. The validity of the algorithm is confirmed by both synthesized signals and a real-world flexible microphone array application.

Speaker Activity Detection and Minimum Variance Beamforming for Source Separation

“…The speakers are assumed to be intermittent which is not the case in most realistic scenarios where source separation has to be applied. Another method proposed recently uses generalized cross-correlation with phase transform (GCC-PHAT) and time difference of arrival (TDOA) together with a clustering algorithm to estimate steering vectors for the speakers, allowing for estimation of speaker activity patterns [9]. However, they did not apply their solution to beamforming for source separation.…”

Latent variable approach to diarization of audio recordings using ad-hoc randomly placed mobile devices

“…In this paper, we consider the diarization of audio recordings using spatial features alone. Several solutions have been proposed utilizing spatial features, which use the time-difference-of-arrival (TDOA) features [4,5,6,7]. However, the estimation of TDOA is sensitive to reverberation and noise.…”