Analysis of Detected Silent Segments in Call Center Recordings

Interpreting speech signals by making a speech to text translation is an active research area especially in current machine learning/deep learning literature. The speech to text translation of call center recordings is an important and specialized application for speech to text translation. Detecting silence in audio recordings can be a pre-processing step in order to optimize processing speed by not-considering audio parts not having significant information. In this work, such a pre-processing framework for detecting silence parts in an audio signal is considered. It is shown that further statistical analysis on the silence distributions results in detecting interesting audio features which can help in finding audio recordings which do not have actual speech sound but a fax machine tone sequence. This foundation can be directly implemented in a call center management software and makes it possible to discriminate between a normal conversation recording and a fax sound recording.