What it does

This is a PyTorch implementation of Listen, Attend and Spell (LAS) and friends: seq2seq attention-based ASR, pure CTC, and joint CTC-attention hybrids. It handles the full pipeline from on-the-fly feature extraction through training to beam-search decoding with optional RNN language model fusion. Everything is driven by YAML config files, with TensorBoard logging baked in.

The interesting bit

The authors don’t pretend this is a turnkey product. The README explicitly warns that example configs “were not fine-tuned,” that TensorBoard error rates are biased (see issue #10), and that CuDNN CTC is “unstable.” That honesty is refreshing in a field of overclaimed benchmarks. The project also serves as a reference implementation for multiple papers, with citations requested rather than demanded.

Key highlights

• On-the-fly feature extraction via torchaudio; character, subword (sentencepiece), or word encoding
• Seq2seq, CTC, and joint CTC-attention training with yaml-driven hyperparameters
• Beam search, greedy decoding, and joint decoding with external RNN-LM
• TensorBoard visualization including attention alignment plots
• Direct lineage from ESPnet and consultation with William Chan (first LAS author)

Caveats

• Decoding is single-sample (no batching), so RAM usage scales with worker count
• Several items remain on the TODO list: pure CTC training has a beam decode bug, examples are “coming soon,” and CLM migration is unfinished
• The project appears to be in maintenance mode; last significant activity unclear from README alone

Verdict

Worth a look for researchers or students who want to understand LAS/CTC internals and don’t mind tuning configs themselves. Skip it if you need a battle-tested, production-ready ASR pipeline—this is a teaching and experimentation scaffold, not a shipped product.