sherpa

Why we included this project

Sherpa is a deployment-focused speech-to-text framework that runs pre-trained end-to-end transducer and CTC models, with C++ and Python APIs for both offline and streaming transcription. It handles the hard parts of putting ASR into production, such as WebSocket server wiring and endpointing, so you don't have to build that plumbing yourself. Since it does inference only, you pair it with models exported from icefall, the companion training toolkit, rather than training anything here. Teams building live captioning, voice assistants, call analytics, or other audio pipelines can quickly stand up a working recognition backend without writing the socket and orchestration plumbing from scratch.