dsh-audio-copilot

Text-only LLMs have no ears and no mouth — users are stuck typing every message character by character. dsh-audio-copilot adds a microphone button right next to the chat input: speak into it, and your words are transcribed into the text box in real time, ready to send.

The core capability of the plugin is speech-to-text, backed by four switchable engines. Zhipu GLM-ASR connects natively inside China and handles Mandarin, Sichuanese, Cantonese, Hokkien, Wu dialects plus dozens of foreign languages at a very low per-use cost. Local faster-whisper runs entirely offline, is free and unlimited, and keeps audio on your own machine. Gemini targets overseas networks and delivers the strongest semantic understanding. Any OpenAI-compatible /audio/transcriptions endpoint is also supported, so existing third-party ASR services plug right in. Switching engines is a one-line config change followed by a restart — nothing else is affected. Reliability details include automatic retry on Gemini 429 rate limits, server-side WebM-to-WAV conversion for the Zhipu engine, actionable Chinese error messages when a key or endpoint is missing, and a clipboard fallback whenever transcription output cannot be injected into the input box, so the plugin never fails silently.

Who is it for? If you are on a Chinese network and need dialect or foreign-language recognition without a proxy, the Zhipu engine is the lowest-friction choice. Developers who value zero cost and offline privacy will prefer local Whisper. Overseas users with a proxy who want the best semantic understanding can pick Gemini. Teams that already run a third-party ASR service can point the plugin at their own OpenAI-compatible endpoint. In every scenario, after installing and restarting DSH, a microphone button appears in the chat toolbar — just talk, and your text is ready.