Shared publicly  - 
 
+srikanth ronanki improved edit distance scoring accuracy from 29% to 73% in less than a week.
3
1
James Salsman's profile photoPat Gunn's profile photo
6 comments
 
+James Salsman It sounds like you have a quantitative approach to improving accent reduction? How well does it, in your experience, actually line up with comprehensibility? It sounds pretty interesting; are your methods English-specific?
 
+Pat Gunn yes; 85% agreement with the per-phoneme judgments of native speakers and expert phonologists is possible (or 80% for both, roughly) is achievable given sufficient exemplar recordings, with much closer agreement on per-word scores, ignoring dialect issues which often confound such agreement; no, but we don't have a pitch tracker for tonal languages yet.
 
+Pat Gunn the duration scoring is solely a function of the exemplar recordings.  If all 30 speak slowly, then it will really hurt the scores of fast talkers. Reflections on whether this is reasonable are very welcome.
 
I'll have to read more about this; does your website have most of the details of the program as a whole and recent innovations, or should I scout around?
Add a comment...