Improving speech detection robustness for wireless speech recognition

The use of speech recognition systems shows that noise and channel effects are very disturbing, and an efficient detection of speech/non-speech segments is necessary. Preprocessing the speech signal is one of the adopted solutions to improve recognition performance. In this paper, spectral subtraction is used as a preprocessing technique aiming to increase the robustness to noisy conditions. Results of several experiments carried out on a database collected over a GSM network show that spectral subtraction improves the global recognizer performance, especially in very noisy environments. We show that the improvements concern mainly noise/speech detection modules