Overlapped Speech Detection in Multi-Party Meetings

Detection of simultaneous speech in meeting recordings is a difficult problem due both to the complexity of the meeting itself and the environment surrounding it. The system proposes the use of gammatone-like spectrogram-based linear predictor coefficients on distant microphone channel data for overlap detection functions. The framework utilized the Augmented Multiparty Interaction (AMI) conference corpus to assess model performance. The proposed system offers enhancements over base line feature set models for classification