SenseVoice
Domain: Models & Providers
Motivation
@deepseek-ai/dsh-experimental-speech-to-text-sensevoice is the local SenseVoice provider: it recognizes speech with SenseVoiceSmall ONNX and Silero VAD on the host CPU, needing no Python, compiler, or model conversion. Preparation downloads revision-pinned, size- and SHA-256-verified model files, and precision defaults to int8 (with fp32 for the larger reference weights).
Concrete Example
The voice bundle supplies an absolute dataRoot under the DSH home and selects sensevoice-local as the default recognizer; the platform-specific sherpa-onnx Node package includes ONNX Runtime.
Analogy
A local transcription engine that runs on your own CPU with no cloud round-trip.