Learn More
• Implemented corrective training to improve recognition performance; on the standard training set this improves speaker-independent perplexity 60 performance from 6.7% error to 5.1% error, and for a(More)