Repository navigation
Releases
Releases · 18301301961/dsh-voice-live
v1.1.0 — live captions + model-assisted correction
Live dictation for the DSH Web composer — words appear while you speak, and the finished transcript lands in the draft.
Features
- Streaming captions: the current sentence is re-decoded on a sliding window every ~0.5–0.8 s
- Sentence-level final decode after Silero VAD endpointing, so final quality matches whole-clip decoding
- 16 kHz PCM captured straight from the microphone (browser echo cancellation / noise suppression / AGC disabled), ITN and punctuation enabled
- Optional model-assisted correction for homophones, product names, digits and punctuation — minimal-edit guard, and it silently falls back to the local transcript on timeout or implausible output
- First run prepares the SenseVoice model through DSH's own speech service
- Audio never leaves the machine
Install
Not published on npm yet. Clone this repository and install the directory as a DSH bundle (see PUBLISH.md), or wait for the npm release.
Tested with DSH Desktop 0.2.2-rc.1 on Windows x64.