Recognition from GSM digital speech

This paper addresses the problem of speech recognition in the GSM environment. In this context, new sources of distortion, such as transmission errors or speech coding itself, significantly degrade the performance of speech recognizers. While conventional approaches deal with these types of distortion after decoding speech, we propose to recognize from the digital speech representation of GSM. In particular, our work focuses on the 13 kbit/s RPE-LTP GSM standard speech coder. In order to test our recognizer we have compared it to a conventional recognizer in several simulated situations, which allow us to gain insight into more practical ones. Specifically, besides recognizing from clean digital speech and evaluating the influence of speech coding distortion, the proposed recognizer is faced with speech degraded by random errors, burst errors and frame substitutions. The results are very encouraging: the worse the transmission conditions are, the more recognizing from digital speech outperforms the conventional approach.