Token Map Drafting speeds up transformer ASR decoding on CPU by using a precomputed n-gram token map as a model-free draft for speculative decoding.
SpeechT5: Unified- Modal Encoder-Decoder Pre-Training for Spoken Language Processing,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
REJECT 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
Model-free Speculative Decoding for Transformer-based ASR with Token Map Drafting
Token Map Drafting speeds up transformer ASR decoding on CPU by using a precomputed n-gram token map as a model-free draft for speculative decoding.