dsh-asr-voice

You speak a sentence into the mic and get back a garbled, run-on mess that still needs a minute of editing before it makes sense as a prompt. dsh-asr-voice, a DeepSeek Harness client-side plugin, closes that gap: transcribed speech is cleaned into a ready-to-send draft in about a second, and an LLM polishes it in the background, swapping in the refined version only when it is done so your manual edits are never overwritten.

The recognition engine is hybrid. Browser Web Speech API takes priority (free, no key needed on Chrome and Edge), and if it is unavailable or blocked the plugin falls back to cloud ASR with multiple providers that can coexist – OpenAI, Groq, SiliconFlow, Xiaomi MiMo, and Alibaba Qwen-ASR. The settings page is a three-step wizard: pick the engine, click a provider, check the key status. Users who already have the same provider configured as an LLM credential in DSH typically never type a key at all. The realtime voice-chat mode turns the interaction into a conversation: live captions appear as you speak, a short pause fires the agent turn, the reply is read back sentence by sentence, and the mic hands itself back for the next utterance. Half-duplex gating prevents echo during playback, and three independent affordances – button, hotkey, or the inline dismiss – let you interrupt at any moment.

If you think faster out loud than you can type, juggle several ASR or LLM providers, or simply want a voice-input experience close to Codex without being locked into a single vendor, dsh-asr-voice is a lightweight option with zero third-party runtime dependencies. It can run alone or sit alongside other DSH plugins. On the privacy side, API keys live exclusively in the DSH credential store; the browser only ever shows a configured or not-configured badge, and audio is transcoded locally before being proxied through a private JSON endpoint.