Acoustic-phonetic features necessary for rule-based speech recognition are described. The selection of feature algorithms is based on their robustness to variability of dynamics in speech signals and to influence of environmental noise, and their suitability for real-time implementation using speech production or speech perception modelling. Features are estimated in real-time, using a 32-bit floating point signal processor DSP32. The features described are pitch, formants, segmentation and labelling. The paper indicates the potential use of these features in connection with future experiments leading to development of new hearing aids.
|Book series||Acta oto-laryngologica. Supplementum.|
|Publication status||Published - 1990|