Automatic Detection of Shouted Speech Segments in Indian News Debates

Shouted speech detection is an essential pre-processing step in conventional speech processing systems such as speech and speaker recognition, speaker diarization, and others. Excitation source plays an important role in shouted speech production. This work explores feature computed from the Integrated Linear Prediction Residual (ILPR) signal for shouted speech detection in Indian news debates. The log spectrogram of ILPR signal provides time-frequency characteristics of excitation source signal. The proposed shouted speech detection system is deep network with CNN-based autoencoder and attention-based classifier sub-modules. The Autoencoder sub-network aids the classifier in learning discriminative deep embeddings for better classification. The proposed classifier is equipped with attention mechanism and Bidirectional Gated Recurrent Units. Classification results show that the proposed system with excitation feature performs better than baseline log spectrogram computed from the pre-emphasized speech signal. A score-level fusion of the classifiers trained on the source feature and the baseline feature provides the best performance. The performance of the proposed shouted speech detection is also evaluated at various speech segment durations.