Online meeting recognizer with multichannel speaker diarization
Shoko Araki, Takaaki Hori, Masakiyo Fujimoto, Shinji Watanabe, Takuya Yoshioka, Tomohiro Nakatani, Atsushi Nakamura · 2010
We present our newly developed real-time conversation analyzer for group meetings. The goal of the system is to estimate automatically “who speaks when and what” in an online manner. In our system, “who speaks when” information is first obtained by estimating the directions of arrival (DOAs) of signals. Then, “who speaks what” is estimated with our automatic speech recognition (ASR) system, after suppressing reverberation, background noise, and interference speakers' voices. In this paper, we focus particularly on the speaker diarization (“who speaks when” estimation) method, and we show that the speaker diarization information helps the ASR to reduce insertion errors.