kaituoxu/Listen-Attend-Spell

The implementation uses a encoder-decoder architecture with attention mechanisms, where the listener (encoder) processes acoustic features via bidirectional LSTMs and the speller (decoder) generates character sequences autoregressively. It integrates with Kaldi for acoustic feature extraction and includes built-in support for Visdom visualization of training metrics. The framework includes a complete pipeline for data preparation, training, and decoding, with example scripts targeting the AISHELL Mandarin speech corpus.

207 stars. No commits in the last 6 months.