The Effects of Speech on Speech Recognition and Tex

Speech is a natural input/output modality for wireless access to information on the web. One way to overcome resource constraints on current wireless devices is to locate the speech recognition and text-to-speech systems on remote servers and transmit compressed speech between the server and wireless clients. To this end, we evaluated the effects of speech compression on the performance of a continuous speech recognizer (with and without noise compensation) and a commercially available text-to-speech system. Standard speech coders such as G.729 (8.0 kb/s), GSM EFR (12.2 kb/s), GSM FR(13.0 kb/s), and G.726 (32.0 kb/s) were used. Our results show that for recognition of digit strings and command phrases recorded in a variety of noise conditions, degradation due to speech coding is more pronounced for clean speech than noisy speech. The GSM coders actually improved recognition performance on noisy speech. For text-to-speech, the GSM EFR was ranked the highest in our A/B listening tests.