labeling
| Playback and analysis by a human operator of speech recordings during a call to a speech-based application (speech traces). For each word pronounced, the pronunciation is checked and a literal or phonetic transcription is made and the quality is examined (correct, noisy, truncated, hesitation).
If possible, the gender and geographical origins of the speakers are also checked. Labeling is used to monitor the behavior of users in relation to the service and to detect any problem areas with the dialog, in order to improve the overall usability of the service. Recommendations for prompt tuning, additionnal grammar development and/or performance enhancement via "doping" of the recognition models could result from this analysis.
| |
| |
