dsh-voice-mimo
Xiaomi MiMo-powered voice for DeepSeek Harness: a browser-native 🎤/🧠/🔊 UI
plus agent tools — voice_transcribe (MiMo ASR), voice_understand
(MiMo-v2.5 semantic analysis), voice_speak (MiMo TTS with a configurable
voice map: preset / voicedesign / voiceclone).
Fork of zhuiyueya/dsh-voice (MIT).
The Settings page structure follows Anionex/dsh-vision-toolkit (MIT).
All three copyright notices are preserved in LICENSE.
What it gives you
| Layer | Capability | Backend |
|---|
| 🎤 | Voice input — mic button in the composer, transcript written into the input box | Browser Web Speech API (zero key) |
| 🔊 | Read-aloud — speaker button on every assistant reply, voice configurable in Settings (朗读音色) | Xiaomi MiMo TTS via host /_dsh/voice-mimo/speak |
| 📄 | voice_transcribe tool — audio file → text | Xiaomi MiMo ASR (mimo-v2.5-asr) |
| 🗣️ | voice_speak tool — text → spoken audio file; renders as a playable strip/card in the conversation | Xiaomi MiMo TTS (mimo-v2.5-tts / -voicedesign) |
| ⚙️ | Settings page — 朗读音色 for 🔊 + voice map + audio.inlineThreshold/longRetain* retention policy, live-applied | DSH Settings (vision-toolkit pattern) |
Unlike upstream dsh-voice (which targets OpenAI-compatible /audio/transcriptions
and /audio/speech endpoints), this fork calls the MiMo API directly — MiMo has
no OpenAI-compatible audio endpoints, so the tools are wired to its native
chat-completions format (ASR text in the assistant message, TTS in audio).
Audio output routes (host)
The 🔊 read-aloud path runs entirely through two same-origin host routes
(node fetch/fs — no 64KB shell stdout cap):
POST /_dsh/voice-mimo/speak {text} — synthesize via MiMo TTS into
audioDir/tmp/ (default ~/.dsh/cache/voice-mimo/tmp/), record a manifest
entry, return {id, audioUrl, bytes, voice, model}. The voice comes from
Settings tts.voice (朗读音色) at request time, so a Settings change
applies on the next click.GET /_dsh/voice-mimo/audio/<id>.wav — stream a stored file to the browser
(id resolved through the manifest, path confined to the audioDir subtree).
Storage skeleton (layered tmp/ + long/, per the audio-output spec):
audioDir (Settings `audio.dir`, default ~/.dsh/cache/voice-mimo/)
├── tmp/ 🔊 read-aloud artifacts — play-once; cleared on DSH startup
├── long/ agent voice_speak artifacts — playable/downloadable strips
└── manifest.json append-only JSONL: {id, sessionId, callId, path, createdAt, text, voice, model, notify}
The plugin reads/writes only inside its audioDir subtree; voice_speak still
respects an explicit outPath. DSH startup (apply) recreates the skeleton
and clears leftover tmp/ contents idempotently.
Agent speech in the conversation (#3)
voice_speak without an explicit outPath:
- writes the wav into
audioDir/long/ and appends a manifest row carrying the
calling session id + call id (for later archive cleanup / regenerate);
- returns
{path, bytes, audioUrl, seconds, notify} — the audioUrl streams
the file via GET /_dsh/voice-mimo/audio/<id>.wav;
- the client renders the tool result as a compact play strip (≤
audio.inlineThreshold seconds, default 30) or a full card (> threshold),
each with ▶ playback and ⬇ download.
With an explicit outPath the exact path is written instead (no strip, no
manifest row). The Settings audio.* fields (inlineThreshold,
longRetainCount, longRetainDays) tune presentation/retention live.
Speaking style & singing (#6/#7)
MiMo TTS exposes rich prosody control; voice_speak surfaces it as:
style (default Settings tts.style = "温柔"): a natural-language
instruction (e.g. 轻快上扬, or a full director-style paragraph). The style
is applied through a mixed channel — preset/voiceclone voices carry it in the
user message; voicedesign voices (whose user message is the voice description)
get an inline (style) tag prefix on the text.sing: true (preset voices only): the text is prefixed with (唱歌) —
put lyrics in text. Combined with a style it becomes (唱歌 style).
Inline tags in the text ((风格), [叹气], …) pass through untouched.truncated: text beyond 2500 chars (official segmentation guidance) is
cut explicitly and the result flags truncated: true.
The manifest row records style so a later regenerate can replay the timbre
plus the prosody. Preset voices follow the official list (mimo_default /
冰糖 / 茉莉 / 苏打 / 白桦 / Mia / Chloe / Milo / Dean).
Install
dsh plugin --profile web add /path/to/dsh-voice-mimo
# or: dsh plugin --profile web add github:ch1bug/dsh-voice-mimo
Configure the MiMo key through DSH Credentials as XIAOMI_API_KEY (the web
Models page writes it), then open Settings → Voice to adjust the voice map.
Pairing: audio file input
Drag-and-drop / paste of audio files into the workspace path is provided by
the separate dsh-drop-to-path
plugin — install it alongside so audio files reach the agent as workspace
paths that voice_transcribe can read:
dsh plugin --profile web add /path/to/dsh-drop-to-path
License
MIT — see LICENSE. Upstream dsh-voice (zhuiyueya) and the
vision-toolkit settings pattern (Anionex) retain their copyright notices.
MiMo API 调用规范(官方文档确认)
- 音频理解(
mimo-v2.5):messages 可带 system 身份提示;content 用
input_audio(data URL,Base64 ≤50MB)+ text 提示词;格式 wav/mp3/flac/m4a/ogg。
- 语音识别(
mimo-v2.5-asr):仅 wav/mp3,Base64 ≤10MB;content 只含
input_audio(不能带 text part);语种走 asr_options.language
(auto/zh/en)。官方示例与本插件实现一致。
- 语音合成(
mimo-v2.5-tts 系):目标文本在 assistant 消息,音色在
audio.voice;voicedesign 不能带 voice 需 optimize_text_preview。
