dsh-voice-mimo

Wiring voice into DeepSeek Harness hits a wall when the backend is Xiaomi MiMo: MiMo exposes no OpenAI-compatible /audio endpoints, so generic voice plugins that assume /audio/speech or /audio/transcriptions simply cannot connect. dsh-voice-mimo was built to close that gap — it skips the compatibility layer entirely and speaks MiMo's native chat-completions format, calling mimo-v2.5-asr for transcription and mimo-v2.5-tts for synthesis, so voice fits naturally into Harness's conversation loop.

The plugin covers the full input-to-output chain. On the input side, a browser-native mic button (Web Speech API, zero keys) drops transcripts straight into the composer. On the output side, every assistant reply carries a read-aloud button that streams through same-origin host routes, sidestepping Node's 64 KB stdout cap. At the agent-tool level, voice_transcribe turns dropped audio files into text the model can reason over, while voice_speak renders text as a playable strip or card, with natural-language style instructions, a singing mode, and explicit truncation beyond 2 500 characters. Three voice channels — preset, voicedesign, and voiceclone — are all supported, and the Settings page lets you swap the read-aloud voice, remap voices, and tune audio retention in real time.

This is the plugin to reach for if you are building Chinese-language voice workflows in DeepSeek Harness: letting the agent read and reason about audio files, adding one-click read-aloud to replies, controlling prosody with free-text style directives, or experimenting with voice cloning and custom voice design. Install it alongside dsh-drop-to-path for audio-file pairing, set XIAOMI_API_KEY in Credentials, and start speaking.