Audio diarization for turkish broadcast news transcription

In this study, speaker independent acoustic models are adapted for each detected speaker using the output of an audio diarization system. It is observed in the experiments that the word error rate for speech recognition is significantly reduced when speaker adaptation is performed. Also, a new speech source separation algorithm is proposed in order to detect overlapping speech segments. It is shown experimentally that the proposed algorithm yields lower diarization error rates than the case when no speech source separation is applied.