hirofumi0810/neural_sp

Supports multiple encoder architectures (RNN, Transformer, Conformer, TDS convolution) with streaming variants using monotonic attention mechanisms (MoChA, MMA) for low-latency decoding. Integrates with Kaldi for feature extraction and implements hybrid training strategies including CTC/attention, hierarchical objectives, and language model fusion (shallow/cold/deep fusion). Handles diverse output units (phoneme, grapheme, BPE, word) via multi-task learning across 10+ standard ASR corpora including Librispeech, Switchboard, and AISHELL.

594 stars. No commits in the last 6 months.