A first-pass decoding strategy for LSTM language models in LVCSR recombines hypotheses on the last two words before lattice rescoring to achieve competitive results on Hub5'00 and Librispeech with better-than-real-time runtime.
Cache based recurr ent neural network language model inference for first pass speec h recognition,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
eess.AS 1years
2019 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
LSTM Language Models for LVCSR in First-Pass Decoding and Lattice-Rescoring
A first-pass decoding strategy for LSTM language models in LVCSR recombines hypotheses on the last two words before lattice rescoring to achieve competitive results on Hub5'00 and Librispeech with better-than-real-time runtime.