srvk/eesen

Eesen streamlines ASR by training a single bidirectional LSTM to directly map speech to text using Connectionist Temporal Classification (CTC), eliminating HMMs, GMMs, and phonetic decision trees. It offers dual decoding paths: WFST-based decoding that incorporates lexicons and language models, or lexicon-free RNN language model decoding via TensorFlow. Built on Kaldi conventions with GPU-accelerated training and parallel utterance processing, it includes complete example recipes for both phoneme and character-based end-to-end systems.

834 stars. No commits in the last 6 months.