ML Paper Challenge Day 17 — Fast and Accurate Recurrent Neural Network Acoustic Models for Speech Recognition

Papers with Code - Fast and Accurate Recurrent Neural Network Acoustic Models for Speech…

Day 17: 2020.04.28
Paper: Fast and Accurate Recurrent Neural Network Acoustic Models for Speech Recognition
Category: Model/Deep Learning/Speech Recognition

New Concept:

Sequence Discriminative Training

• state-level minimum Bayes risk (sMBR) sequence discriminative training criterion
• sMBR training can fix scaling issue if we do not scale the blank label posterior while decoding an utterance to get numerator and denominator lattices
• blank label scaling can be baked into into the bias of the blank label output unit in the RNN model by adding negative log of the scale before starting sMBR training

Acoustic Features

• reduce the number of input frames
• decimating the input frames…