Using a priori information for speaker diarization

This paper presents an attempt to use supplementary information for audio data diarization. The approach is based on the use of a priori information about the speakers involved in dialogue. Those specific information are the number of speakers involved in conversation, and training data available for one speaker or for all the speakers involved in conversation. The experiments were mainly conducted on the 2003 Rich Transcription Diarization corpus both Dry Run Corpus and Evaluation corpus. The results show that knowing a priori the exact number of speakers seems not to be a very useful information. On the other hand, using a priori speaker models for one or all speakers involved in the conversation, may improve diarization performance when enough data is available to train reliable speaker models.