Iris Merkus
2012
Statistical Evaluation of Pronunciation Encoding
Iris Merkus
|
Florian Schiel
Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12)
In this study we investigate the idea to automatically evaluate newly created pronunciation encodings for being correct or containing a potential error. Using a cascaded triphone detector and phonotactical n-gram modeling with an optimal Bayesian threshold we classify unknown pronunciation transcripts into the classes 'probably faulty' or 'probably correct'. Transcripts tagged 'probably faulty' are forwarded to a manual inspection performed by an expert, while encodings tagged 'probably correct' are passed without further inspection. An evaluation of the new method on the German PHONOLEX lexical resource shows that with a tolerable error margin of approximately 3% faulty transcriptions a major reduction in work effort during the production of a new lexical resource can be achieved.