next up previous contents index
Next: Response time Up: Speech recognition systems Previous: Error measure presentation

Error handling

Different types of error may occur. Some of them may be handled at the application level, others may/should be handled at the speech recognition level or at least forwarded to the application.

Time out

What happens if the speech recogniser  incorporates a speech detection activity and does not detect any speech signal during a period of time? This may be due to the user who does not speak or may be due to the speech detector. So this can be handled at the dialogue level as a time out parameter or it can be used by the recogniser to set up (or adapt) its thresholds of speech end-point detection.


Signal-to-noise ratio

The system may have some difficulties in extracting speech data from a continuum of speech with a low signal-to-noise ratio. If this is detected the application may ask the user to speak louder or use another handset. Some parameters may need a particular tuning (e.g. energy thresholds ) to avoid significant performance degradation with a change of microphone  or telecommunication channel .

For example if the system has been optimised for adverse   and severe conditions and is being used in benign or favourable conditions (or vice versa) this may lead to significant degradation of performance and the application developer should be informed. He may require some figures about the performance obtained with different signal-to-noise ratio levels.  

EAGLES SWLG SoftEdition, May 1997. Get the book...