eesen

Why we included this project

Eesen is a research toolkit that treats speech recognition as a single sequence learning problem. Instead of the HMM, GMM, and decision-tree machinery of a classic pipeline, you train one bidirectional LSTM with connectionist temporal classification (CTC) to map speech directly to text. The two decoding paths are the part worth studying: WFST-based decoding lets you bring lexicons and language models into CTC output efficiently, while the RNN-LM path works without a fixed lexicon. The repo ships full Kaldi-style example setups for phoneme and character labels, so you can reproduce an end-to-end system and study the recipes rather than just read about them. Treat it as a learning base for CTC-based ASR, not a drop-in production service; the project has been dormant since 2019.