MASR

Why we included this project

MASR is a PyTorch speech-to-text framework that handles both streaming and non-streaming recognition, so it fits teams working on live captioning as well as offline transcription. The main models, Conformer, Squeezeformer, and DeepSpeech2, sit behind a single config-driven interface, which means you can swap models and decoders without rewriting your application code. Pretrained checkpoints cover Mandarin, Cantonese, English, and mixed Chinese-English data, giving you a working baseline before you fine-tune on your own audio. It also bundles data augmentation, speaker-separated inference, and a web demo, so a small team can get a self-contained ASR stack running without assembling pieces from scratch.