Spotting chief speaker from press conference recordings based on silence detection

This paper presents a method for spotting chief speaker from press conference recordings where durations of silence segment in one utterance of chief speaker and other speakers are obviously different. In the proposed method, speech endpoint detection is first performed on the audio conference recordings for obtaining the durations of silence segment (i.e. S_i sequence). Then, S_i sequence is converted into “1-0” sequence where outliers are revised. Finally, speech segments limited by the continuous “1” sequences are extracted as the chief speaker´s voices. The experiments are conducted on two data sets with different durations of silence segment in chief speaker´s utterances for comparing the proposed method with the conventional approach (based on speaker segmentation using BIC and spectrum clustering). The experimental results show that the proposed method achieves higher F measures (harmonic mean of precision rate and recall rate) with faster speed in comparison with the conventional approach for spotting chief speaker from press conference recordings.