Urgent Voicemail Detection Focused on Long-term Temporal Variation

This paper proposes a effective urgent speech detection for voicemails focused on speech rhythm. Previous techniques use short-term features with millisecond scale (such as fundamental frequency, loudness and spectral features), and conventional techniques for urgent speech detection use also features obtained from entire speech (such as average speech rate). However, the features obtained from entire speech are too over-smoothed to explain the difference between urgent and nonurgent speech. We found that there was a difference between urgent and non-urgent speech in temporal variability related to speech rhythm. To handle the temporal variability of speech rhythm, the proposal extracts long-term temporal features. The long-term temporal features are envelope modulation spectrum and temporal statistics of Mel-frequency cepstrum coefficient with 1 sec scale. To use both features with different time scales, the proposed method integrates the long-term temporal features and the short-term features on neural networks. Our proposal yields better accuracy than the conventional methods (which uses e features obtained from entire speech); it achieves a 50.0% reduction in the error rate.