dsh-mimo-tts

MiMo voice for DeepSeek Harness — read every answer aloud, talk to the composer, switch voices, and clone your own from a short recording. 中文说明

The plugin speaks through Xiaomi MiMo TTS and listens through MiMo ASR, both on the token-plan endpoint, with the same key your LLM providers already use. Nothing is uploaded anywhere except to Xiaomi's own API.

What it does

🔊 Auto-read	Every finished answer is spoken. Code fences, tables, links and markdown scaffolding are stripped first, and long answers are clipped so a wall of text cannot turn into three minutes of speech.
🎙 Click to talk	A mic button sits in the composer's own control row (next to the model picker). Click to start, click again — or simply stop talking and it ends itself on the pause — and the transcript lands in the input box, ready for Enter.
⏹ Stop	The moment she starts speaking, a stop button appears next to the mic; Esc does the same. Stopping also drops the sentences still queued behind the current one, and opening the mic cuts her off automatically.
🗣 Voices	Nine built-in MiMo voices, plus a cloned voice slot driven by a local 5–15 s reference sample.
🎛 Panel	Plugins → MiMo voice (and Settings → Built-in plugins): pick a voice and hear it immediately, toggle auto-read, type something to test, re-read the last answer, watch queue/status.

Requirements

• DeepSeek Harness 0.2.0-rc.1 or newer (engines.dsh)
• A Xiaomi MiMo token-plan key, reachable either from
  • the XIAOMI_TOKEN_PLAN_CN_API_KEY environment variable, or
  • refs: in $DSH_HOME/.credentials.yaml — the same place the LLM providers read.
• The key is only sent to token-plan-cn.xiaomimimo.com. Note that a token-plan key is not accepted by the public api.xiaomimimo.com host, and vice versa.

Install

dsh plugin --profile web add dsh-mimo-tts      # or --profile desktop
# restart the host (dsh web / the desktop app)

Use

1. Auto-read: on by default after you switch it on once from the panel. Every finished answer is queued and played in order.
2. Talk: click the mic in the composer, speak, pause (or click again). The text appears in the input box — press Enter to send.
3. Stop: click the red stop button, or press Esc.

Voices and cloning

Nine built-in voices: mimo_default, 冰糖, 茉莉, 苏打, 白桦, Mia, Chloe, Milo, Dean.

To use your own voice, drop a 5–15 second, clean, single-speaker clip at

$DSH_HOME/.dsh-mimo-tts/voice-clone.mp3      # mp3 or wav

and select 「克隆音色」/ “Cloned voice” in the panel. The host reads that file directly; it is never bundled with the plugin and never leaves your machine except as the per-request reference the MiMo API requires.

Cloning costs latency: every request carries the reference sample (~105 KB for a 10 s clip), so a sentence takes ~3 s instead of ~1 s. Please only clone a voice you have the right to use.

Configuration

$DSH_HOME/.dsh-mimo-tts/config.json (the panel writes this file):

{
  "voice": "冰糖",       // any built-in voice, or the clone slot
  "autoRead": true,      // speak every finished answer
  "maxChars": 600,       // clip long answers before speaking
  "format": "mp3"        // mp3 | wav
}

HTTP surface

All routes live under /dsh-mimo-tts on the host's own webserver:

Route	Purpose
GET /status	key presence, model names, voice roster, clone sample state, queue depth
GET /voices	the voice roster
POST /config	{ voice?, autoRead?, maxChars? }
POST /speak	{ text, voice?, queue? } → synthesized audio as a data URL
POST /speak-last	speak the last finished answer of a session
POST /asr	raw audio/wav (16 kHz mono) or audio/mpeg → { text }
GET /pending?since=N	drain the auto-read queue (cursor based, nothing repeats)
POST /stop	stop now: the client pauses, the host drops the queue

Security: requests carrying a cross-site Origin are refused with 403; loopback, dsh-app://, file:// and origin-less callers are allowed. DSH_MIMO_TTS_ORIGINS adds origins for deployments that serve the GUI from another host.

Development

node test/smoke.mjs                 # offline: routes, fence, sanitizer, validation
DSH_MIMO_TTS_LIVE=1 node test/smoke.mjs   # adds one real synthesis + transcription

The browser half (lib/client.js) is a hand-written __ModuleLoader__ bundle: no build step, React comes from the host's module table.

License

MIT — see LICENSE.