There have been several new systems for multimedia information access reported in recent years. The system presented here shares many of their aspects, but it differs in a significant way from them; it extends the realm of multimedia access to include speaker-based information. We have already prototyped and reported such a system elsewhere whose main features include SVAPI-based speaker recognition combined with speech recognition for joint text-and speaker-based retrieval from audio and

more »

... A vital component of such a system is speaker identification whose performance degrades for utterances smaller than eight seconds to such an extent that such segments have to be dismissed with a catch-all, neutral label. Here, we use a Bayesian Information Criterion based speaker clustering technique to analyze the same audio data. The results of this classifier are combined with those from our SVAPI-based speaker classifier using a decision integration scheme to produce new labels for many such short speaker segments. We discuss the details of this combined analysis and its results. We additionally report on a on-the-fly speaker enrollment scheme using this BIC-based speaker clustering technique.