Usable Speech Detection Using Linear Predictive Analysis - A Model-Based Approach
Nithya Sundaram, Robert E. Yantorno, Brett Y. Smolenski, Ananth N. Iyer, Norris Streets · 2003
A speech segment is defined as usable, if speech, which is corrupted by interfering speech, can still be used for applications like speaker identification. In tactical communications, where there are multiple signals transmitted over the same channel such as telephone or radio transmission, separation of usable speech from speech corrupted by voices of other speakers is desired. This separation is important in making automatic speaker and speech recognition systems more robust. A novel approach towards developing a usable speech measure could be model-based. Using this concept of model-based usable speech detection, the use of Linear Prediction is investigated. The method reveals that an average of 75% of the usable speech is correctly detected with false alarms of 34%.