Speech-to-text inference for the end-to-end purist

What it does sherpa is a speech-to-text inference framework built on PyTorch that exclusively serves end-to-end models—specifically transducer and CTC architectures. It provides C++ and Python APIs and focuses strictly on deployment, not training. For fine-tuning or training, it redirects you to its companion project icefall.

The interesting bit Rather than swelling into a kitchen-sink toolkit, the project acknowledges its PyTorch weight and points to siblings sherpa-onnx and sherpa-ncnn for iOS, Android, and embedded targets. It is a rare case of a project family cleanly partitioned by runtime rather than feature.

Key highlights

• End-to-end only: deliberately ignores classical hybrid pipelines in favor of transducer and CTC models.
• Dual-language APIs: exposes both C++ and Python interfaces for server integration.
• Browser demo: includes a Hugging Face space for zero-installation trials.
• Clear scope boundary: training is explicitly out of scope, with a direct hand-off to icefall.

Caveats

• The README is minimal; substantive documentation lives on an external site.
• PyTorch dependency makes it a poor fit for mobile or deeply embedded use, though the siblings cover that gap.

Verdict Useful if you need a PyTorch-native server for modern ASR inference. Look elsewhere—specifically at its onnx or ncnn siblings—if you need lightweight edge deployment, or at icefall if you need to train.