Lower Frame Rate Neural Network Acoustic Models
Lower Frame Rate Neural Network Acoustic Models
G. Pundak,Tara N. Sainath
2016 · DOI: 10.21437/Interspeech.2016-275
Interspeech · 142 citaten
TLDR
On a large vocabulary Voice Search task, it is shown that with conventional models, one can slow the frame rate to 40ms while improving WER by 3% relative over a CTC-based model, thus improving overall system speed.