Reproduce “Imputer: Sequence Modelling via Imputation and Dynamic Programming” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
William Chan, Chitwan Saharia, Geoffrey Hinton, Mohammad Norouzi, Navdeep Jaitly
Imputer addresses speech recognition’s decoding bottleneck by refining a partially filled alignment in parallel across blocks, committing one prediction per block per iteration. Dynamic programming training sums probabilities of completions compatible with a sampled partial alignment. On LibriSpeech test-other, it scores 11.1 WER, versus 13.0 for CTC and 12.5 for seq2seq; this result does not measure decoding time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.