mimo-voice-report

MiMo Voice Report

English · 中文

A DeepSeek Harness plugin that speaks a short completion notice and a two-sentence summary every time a task finishes — and pops up a portrait with a speech bubble whose text types itself out in step with the audio.

Powered by Xiaomi MiMo TTS. Ships with a 7.7-second reference clip, so the default setup already speaks in a cloned voice (mimo-v2.5-tts-voiceclone); switch one line to use a preset voice instead.

dsh-plugin · deepseek-harness · cordis-plugin · mimo · tts · voice-clone

What it does

turn ends (session log gets turn/end)
   │
   ▼
Host half (index.js)                        Client half (client.js)
· folds the last assistant text +           · speaker button in the composer toolbar
  end reason from session/event             · useSession(s => s.running) true → false
· on POST /mimo-voice/report:               → POST /mimo-voice/report
  1. asks MiMo chat to rewrite the          · plays the base64 mp3
     result into 1–2 spoken sentences       · pops the portrait + speech bubble,
  2. synthesizes it with MiMo TTS              text revealed by audio progress
  3. returns base64 mp3 (cached per turn)

• Completion line: 任务完成啦。 (error / aborted / blocked have their own lines).
• Summary: one MiMo chat call (mimo-v2.6-flash by default) rewrites the last assistant message into one or two speakable sentences; if that call fails or is rejected, a deterministic extractive summary is used instead.
• Persona-aware summaries (v1.20.0): every voice ships a persona prompt that steers the summary model's tone — the tsundere reports "hmph, it's not like I did it for you", the cool pharmacist keeps it clipped and precise, the ojou-sama stays refined. 16 built-in prompts; override or add your own per voice (including custom samples) in the settings card's 人设提示词 editor (config.personaPrompts[id], max 200 chars — describe *who* and *how they speak*; structural rules are appended by the plugin). The emergency extractive fallback bypasses personas.
• No duplicate openings: the model sometimes repeats "任务完成啦"; dedupeAnnounce() strips a leading completion phrase and falls back to the extractive path if nothing is left.
• Pay only when needed: synthesis happens only for the session you are looking at; audio is cached per turn (a cache hit skips even the summary call), and concurrent requests for the same turn (two tabs, rapid clicks, a force retry) are merged into a single generation.
• Key never reaches the browser: the API key is read in the host process; the page only receives mp3 bytes.

Install

git clone https://github.com/387617119/mimo-voice-report.git

Then install the directory as a bundle in your DSH profile:

plugin_manager  action: install_bundle  target: /absolute/path/to/mimo-voice-report

(Any equivalent "install a local bundle directory" flow works — the package declares dsh.bundle.patch.)

Configure

Everything lives in this package's cordis.patch.yml row, and most of it can be edited at runtime in Settings → Plugins → MiMo Voice Report.

Key · Default · Meaning baseUrl · https://token-plan-cn.xiaomimimo.com/v1 · OpenAI-compatible endpoint. Token-plan keys use token-plan-cn; pay-as-you-go keys use https://api.xiaomimimo.com/v1 apiKey · '' · MiMo API key. Empty falls back to MIMO_API_KEY, then ~/.dsh/.credentials.yaml refs model · mimo-v2.5-tts-voiceclone · mimo-v2.5-tts (preset voice), mimo-v2.5-tts-voiceclone (clone a clip), mimo-v2.5-tts-voicedesign (text-designed voice; timbre drifts per call) voice · 冰糖 · Preset voice: 冰糖, 茉莉, 苏打, 白桦, Mia, Chloe, Milo, Dean voiceSample · voice-sample.mp3 · Reference clip for clone mode, relative to the package or absolute; mp3/wav, mono, 5–15 s of clean speech avatar · avatar.webp · Portrait for the popup (png/webp/jpeg/gif); missing/empty = bubble only style · '' · Optional delivery instruction sent as a role: user message (never spoken). Leave empty for the most faithful clone tags · (俏皮 稚嫩) · Audio style tag prepended to the text (playful + girlish-cute). '' disables it; tags can shift timbre callMe · '' · How the whale girl addresses you, appended to the opening line (任务完成啦，主人。). CJK/Latin letters, digits or ·, up to 16 chars; empty = no address customAnnounces · {} · Your own lines per bucket — {completed, error, aborted} (one per line, max 20 × 40 chars each) — shuffled into the matching built-in pool; "interrupted" shares the aborted bucket. Write the address into a line itself if you want it customOnly · false · true = a bucket with custom lines uses only yours, none of its built-ins volume · 1 · Playback volume 0–1 (0–100%). The settings-card slider applies live while dragging; Save writes back here. The browser half applies it to audio.volume personaPrompts · {} · Per-voice persona prompts (voiceId: sentence) steering the summary model's tone; empty uses built-in defaults; the settings-card editor writes back here language · zh · Report language (zh/en/ja/ko). Other languages switch the spoken audio to that language (the whole Chinese report is MiMo-translated, persona tone kept); the bubble subtitle follows the spoken text and appends the Chinese original as a translation once the audio finishes. Falls back to Chinese for that turn if translation fails. Dialect voices only show up under `zh` — switching to another language hides them and falls back to the original voice if one was selected speechRateFactor · '' · Speech-rate multiplier (0.5–2). Empty uses per-language defaults: zh 1.00 / en 1.08 / ja 1.12 / ko 1.10. Switching language automatically drops any custom factor and applies the new language's default (set it again to pin it). Applied after synthesis with ffmpeg atempo (pitch preserved) — prompt/tag-based rate control proved unreliable on MiMo (same Japanese line ranged 9.02–11.74 s) personaPrompts · {} · Per-voice persona prompts; the six dialects ship their own (dialect-yue/-sc/-ne/-ha/-sd/-tw): Cantonese pragmatic warmth, Sichuan hot temper straight talk, Northeastern open-hearted bluster, Henan plain-spoken honesty, Shandong generous down-to-earth reliability, Taiwan soft cutesy — all overridable fallbackDing · fallback-ding.wav · Bundled chime played when MiMo is unavailable (path relative to the package); '' disables the fallback cooldownMs · 300000 · Cooldown after 3 consecutive MiMo failures — the API is skipped entirely; clicking the speaker (force) bypasses it summarizer · model · model = rewrite with MiMo chat; extract = rule-based, one less API call summaryModel · mimo-v2.6-flash · Model used for the rewrite summaryMaxChars · 64 · Max length of the spoken summary maxChars · 90 · Max length when falling back to the extractive summary staleMs · 25000 · Don't speak for a turn that ended longer ago than this timeoutMs · 30000 · Per-request timeout

API key

Three ways, in resolution order:

1. In the UI — Settings → Plugins → *MiMo Voice Report* → paste the key → Save. It is written back to this plugin's row in your profile patch (config.apiKey) and takes effect immediately. Clear removes the override and falls back to the sources below.
2. MIMO_API_KEY environment variable.
3. refs in ~/.dsh/.credentials.yaml — MIMO_API_KEY, then MIMO_TTS_API_KEY.

The settings card shows the active source plus a masked hint (first 3 + last 4 characters). GET /mimo-voice/config never returns the full key.

Voice presets

The repo ships sixteen synthetic reference clips — all rendered by MiMo itself, no original recording is bundled. Switch between them in Settings → Plugins → MiMo Voice Report → 音色, or edit voiceSample in cordis.patch.yml:

Preset · File · Character 原版 (original) · voice-sample.mp3 · Neutral demo take of the test line 俏皮版 (playful) · voice-sample-playful.mp3 · Playful take of the same timbre 傲娇小猫 (tsundere) · voice-sample-tsundere.mp3 · Bright teen voice, brisk, prickly-outside-caring-inside 清冷药师 (kuudere) · voice-sample-kuudere.mp3 · Cool mid-low voice, calm, clipped, deadpan wit 元气应援 (genki) · voice-sample-genki.mp3 · High, cheery, cheerleader energy, rising endings 温柔姐姐 (onee-san) · voice-sample-onee.mp3 · Soft breathy mid-low voice, unhurried, reassuring 病娇低语 (yandere) · voice-sample-yandere.mp3 · Sweet whisper, breathy, slow, clinging undertone 无口人偶 (doll) · voice-sample-doll.mp3 · Flat, quiet, minimal affect, doll-like recital 软萌萝莉 (loli) · voice-sample-loli.mp3 · Very high round childlike voice, milky articulation 涩谷辣妹 (gyaru) · voice-sample-gyaru.mp3 · Loud bright brash gal voice, teasing, drawn-out accents 名门大小姐 (ojou-sama) · voice-sample-ojou.mp3 · Refined elegant lady, precise diction, politely lofty 天然呆 (airhead) · voice-sample-tennen.mp3 · Soft warm voice, relaxed half-a-beat-off rhythm, innocent 小恶魔 (little devil) · voice-sample-koakuma.mp3 · Silky nimble voice, loves to tease, smug rising endings 假小子 (tomboy) · voice-sample-bokukko.mp3 · Mid-low crisp voice, boyish and loyal, no-nonsense 阴角少女 (introvert) · voice-sample-inkyo.mp3 · Light timid voice, small and hesitant yet articulate 文学少女 (bookworm) · voice-sample-bunge.mp3 · Quiet clear voice, read-aloud clarity, bookish calm

The fourteen archetype voices are fully synthetic: mimo-v2.5-tts-voicedesign rendered each from a textual acoustic description of a popular anime-girl voice archetype (tsundere / kuudere / genki / onee-san / yandere / emotionless / loli / gyaru / ojou-sama / airhead / little devil / tomboy / introvert / bookworm). No real person's or voice actress's recording is cloned, and the preset names are original aliases with no reference to any specific character — the descriptions stay at the acoustic-archetype level on purpose. Every sample is screened by objective F0 metrics (median pitch, spread, duration); the third batch onward follows a uniform "natural tone, moderate pace" recipe.

The selector writes back config.voiceSample and takes effect immediately. The 试听 (preview) button next to it plays the bundled fixed persona-line MP3 for the selected voice (preview-.mp3, generated by cloning the matching sample with a persona-matched line — instant, local, zero MiMo usage). To hear custom text, type a line below and press 生成试听 (or Enter) — that synthesizes on the fly via /say (rate-limited to 6/min). Next to the speaker on the conversation page there is a gear button (shown when the host exposes pluginNavigation, dsh 0.2.1+) that jumps straight to this plugin's settings page — voice, volume and key all live there. The settings card also has a volume slider (0–100%): dragging applies live to the next playback; Save persists it to config.volume. Any voice-sample-*.mp3 / .wav you drop next to index.js shows up in the same dropdown (fixed preview optional; without one the button falls back to custom-line generation) — the host only accepts files that actually exist in the package directory.

The two base clips are synthetic: MiMo voice-clone rendered a test line 「test, test, 这里是音频测试, 1, 2, 3, 4, 5, 上山打老虎」 in the author's chosen timbre. To use your own voice:

Goal · How Use your own voice · Overwrite voice-sample.mp3 (or add a new file) with 5–15 s of single-speaker dry speech Use a MiMo preset voice · model: 'mimo-v2.5-tts' + voice: '冰糖' (or 茉莉, 苏打, 白桦, Mia, Chloe, Milo, Dean) Only change the delivery · Put an instruction in style (sent as role: user, never spoken), or a tag in tags

Normalize a new clip with:

ffmpeg -i raw.mp3 -af loudnorm=I=-16:TP=-1.5:LRA=11 -ac 1 -ar 44100 -b:a 128k voice-sample.mp3

The portrait is not bundled (avatar.webp missing → the popup shows the bubble only). Drop your own avatar.png/avatar.webp next to index.js to get the character card.

Notes from building this:

• The clone model wants audio.voice as a DataURL (data:audio/mpeg;base64,...) — a bare base64 string is rejected with HTTP 400.
• Longer clips (15–20 s) are not automatically more faithful; measure your own result. A 5 s clip can give an exact pitch match while a longer one drifts.
• seed is accepted but ignored, and temperature: 0 does not make takes reproducible: the same text can land a few semitones apart between calls. If consistency matters more than likeness, use a preset voice.
• If your reference has background music, isolate the vocal first, e.g. python -m demucs --two-stems=vocals out.mp3 and use vocals.wav.
• The bundled clip is a voice-clone render of a test line, not a recording. Reproduce it with any reference clip you own:

{
    "model": "mimo-v2.5-tts-voiceclone",
    "messages": [{ "role": "assistant", "content": "test，test，这里是音频测试，1，2，3，4，5，上山打老虎" }],
    "audio": { "format": "mp3", "voice": "data:audio/mpeg;base64,<your reference clip>" }
  }

Then normalize the returned audio (loudnorm=I=-16, mono, 44.1 kHz) and save it as voice-sample.mp3.

• Voice cloning carries its own responsibility: only clone voices you have the right to use.

Endpoints (for debugging)

`` is whatever port the DSH web GUI is on this session (it changes on every start).

curl -s http://127.0.0.1:<port>/mimo-voice/state        # host snapshot: version, sample, avatar, counters
                                                        #   + `events`: last 50 diagnostic events (failures,
                                                        #     breaker open/recovered, fallbacks, cache hits)
                                                        #   + `lastError` with the raw HTTP status & body snippet
curl -s http://127.0.0.1:<port>/mimo-voice/config       # which key is in effect (masked)
curl -s http://127.0.0.1:<port>/mimo-voice/_client      # last browser-half report
curl -s "http://127.0.0.1:<port>/mimo-voice/preview?voice=voice-sample-loli.mp3" -o loli.mp3 \
                                                        # fixed persona-line preview (mp3, zero MiMo usage)
curl -s -X POST http://127.0.0.1:<port>/mimo-voice/say \
  -H 'content-type: application/json' \
  --data '{"text":"任务完成啦。这是一次试听。"}'          # custom-line synthesis (base64 mp3; rate-limited: 6/min)

Development

Pure-function unit tests, no network needed:

npm test        # = node test.mjs

Privacy & security

• The API key lives only in the host process; it is never sent to the page or written into logs.
• All routes are served by the local DSH web server and are only reachable from that origin — do not expose the port to a network.
• The plugin stores no conversation content beyond one in-memory record per session (last assistant text, end reason, timestamp), capped at 60 sessions.

License

MIT — see LICENSE.

Voice cloning is powerful and easy to misuse: only clone voices you have the right to use, and say so when you publish the result.

中文说明

一个 DeepSeek Harness 插件：任务完成时用小米 MiMo TTS 播报一句完成提示 + 一两句结果总结，同时弹出立绘和对话气泡，气泡里的文字跟着音频进度逐字出现。

• 音色二选一：MiMo 预置音色开箱即用（冰糖 等 8 个），或自带 5–15 秒干声做 voice clone
• 口播稿由 MiMo 对话模型改写成"一句提示 + 两句总结"，失败时退回规则摘要
• 输入框右侧小喇叭可随时开关；被浏览器拦截自动播放时点一下补播
• MiMo API Key 三种配法：插件配置页粘贴（会写回 profile 的 patch）、MIMO_API_KEY 环境变量、~/.dsh/.credentials.yaml 的 refs；Key 只在宿主进程使用

安装：把本仓库克隆成目录后，用 plugin_manager 的 install_bundle 指向该目录即可。

声音复刻请自行确认授权范围；发布克隆结果时也请说明来源。