mailong25/self-supervised-speech-recognition

Builds accurate speech recognition for low-resource languages through a three-stage pipeline: self-supervised pretraining on unlabeled audio, fine-tuning on minimal labeled data (as little as 1 hour), and n-gram language model integration for beam search decoding. Leverages fairseq's wav2vec 2.0 implementation with cross-lingual transfer initialization and optional KenLM decoding, enabling practical deployment via a simple Python API despite training resource requirements (V100 GPUs).

379 stars. No commits in the last 6 months.