openspeech-team/openspeech

Provides reference implementations of 20+ ASR architectures (DeepSpeech2, RNN-Transducer, ContextNet, etc.) with support for encoder-decoder and transducer-based decoding strategies including beam search. Includes multi-language recipes for English, Mandarin, and Korean with pre-configured datasets (LibriSpeech, AISHELL-1, KsponSpeech) and modular audio processing with SpecAugment, mel-spectrograms, and MFCC feature extraction. Leverages Hydra's hierarchical configuration system to enable hardware-agnostic training across multi-GPU, TPU, and mixed-precision setups without code changes.

718 stars. No commits in the last 6 months.