Summary
Large reductions in SR WER result in small IR improvements
51k vocabulary is sufficient - low OOV rate
Stemmed language models didn’t help
Confidence Measures provide no benefit
- Deleted (missing) words are most critical
Too little data for conclusive experiments