Adaptive robust speech processing based on acoustic noise estimation and classification

The paper presents an adaptive system for speech signal processing in the presence of loud background noise. The validity of the approach is confirmed by implementing a classification system for voiced and unvoiced (V/UV) speech frames. Genetic algorithms were used to select the parameters that offer the best V/UV classification in the presence of 4 different types of background noise and with 5 different SNRs. 20 neural network-based classification systems were then implemented, chosen dynamically frame by frame according to the output of a background noise recognition system and an SNR estimation system. The system was implemented and the tests performed using the TIMIT speech corpus and its phonetic classification. The results were compared with a non-adaptive classification system and the 3 V/UV detectors adopted by three important: LPClO, ITU-T G. 723.1 and ETSI AMR. In all cases the adaptive V/UV classifier clearly outperformed the others, confirming the validity of the adaptive approach