gentaiscool/end2end-asr-pytorch

Implements a low-rank Transformer encoder-decoder architecture with optional CNN feature extractors (VGG or embedding-based) for sequence-to-sequence ASR. Supports multi-GPU batch parallelization and flexible training across multiple datasets with configurable beam-search decoding, achieving 13.5% CER on Mandarin Chinese with beam width 8. Compatible with PyTorch and torchaudio, accepts custom datasets via CSV manifests with character-level labeling.

304 stars. No commits in the last 6 months.