dsh-minimax-voice

Voice cloning, text-to-speech and speech-to-text for DeepSeek Harness, powered by MiniMax.

A community plugin for DeepSeek Harness. Not an official DeepSeek or MiniMax product. A MiniMax account with voice cloning, TTS and (optionally) speech recognition enabled is required.

Clone a voice from a 10-second sample, hear the result without paying the ¥9.9-per-voice cloning fee, then use that voice_id from the Agent, from the settings panel, or from your own plugin. Audio can also be transcribed back into text.

Features

• Voice cloning — mp3 / m4a / wav, 10 s–5 min, ≤20 MB. Optional second sample (<8 s) improves similarity.
• Free preview, deferred billing — the preview comes from the demo_audio returned by /v1/voice_clone, billed as TTS characters only. The ¥9.9 per-voice cloning fee is not triggered; it is charged when you first synthesize with the voice. See Billing.
• Noise reduction toggle — strip sample hiss before cloning (on by default; MiniMax's own default is off).
• Text-to-speech — 10 000 characters per call, eight emotions, speed / volume / pitch control, MiniMax system voices or your own clones.
• Speech-to-text — asr-1.0, ≤500 s / ≤50 MB per call, automatic language detection, mixed Chinese-English handled, returns plain text.
• Four Agent tools — the model can clone, list, synthesize and transcribe on its own.
• Settings panel — key management, cloning form with live preview player, TTS form, ASR card, voice library with delete.
• Local only — settings live in ~/.dsh/minimax-voice/, the HTTP service binds to localhost and is only reachable from the Harness UI.

Why the preview is free

MiniMax charges voice cloning not when /v1/voice_clone is called, but on the first T2A synthesis using that voice_id (¥9.9 per voice). A preview implemented as "clone, then call T2A to hear it" therefore burns the fee immediately.

This plugin instead reads demo_audio from the clone response — the sample the API itself renders — downloads it locally and plays that. The voice is stored with activated: false and marked · 未激活 in the panel. activated flips to true on the first real synthesize(), which is also when MiniMax starts counting the ¥9.9.

Requirements

• DeepSeek Harness >=0.2.0 (Web or desktop build that includes the Web UI).
• Node.js 22+.
• A MiniMax API key with T2A and voice cloning enabled. Speech recognition is optional and billed separately.
• Network access to api.minimax.chat (international) or api.minimax.cn (China).

Installation

Ask an agent

Send this to any agent that can run terminal commands. Replace desktop with your profile name (dsh profiles lists them):

Install the DSH plugin dsh-minimax-voice into my desktop profile:
dsh plugin --profile desktop add dsh-minimax-voice --registry=https://registry.npmjs.org/
Then confirm it loaded by checking that GET http://127.0.0.1:58500/mmvoice/api/status responds,
and tell me where to paste my MiniMax API key.

Install from npm

dsh plugin --profile desktop add dsh-minimax-voice --registry=https://registry.npmjs.org/

Install from source

git clone https://github.com/Sunshine334419520/dsh-minimax-voice.git
dsh plugin --profile desktop add /absolute/path/to/dsh-minimax-voice

Reload

Desktop apps usually pick the new plugin up on their own. If the settings page does not show MiniMax Voice, toggle the plugin off and on in Plugins, or restart the app. A browser refresh alone is not enough for host-side code, but is needed for panel changes.

Configure

Open Settings → MiniMax Voice, paste your API key and save. The key is written to ~/.dsh/minimax-voice/secret.json. You can also set MINIMAX_API_KEY in the environment or config.apiKey in the plugin config; precedence is config.apiKey → MINIMAX_API_KEY → secret.json.

Verify:

curl -s http://127.0.0.1:58500/mmvoice/api/status
# {"configured":true,"keySource":"file","dataDir":"/Users/you/.dsh/minimax-voice",
#  "defaultModel":"speech-2.8-hd","voiceCount":3,"previewMode":"demo_audio"}

Usage

Settings panel

Clone a voice — pick a sample, optionally rename the voice, keep noise reduction on, press 复刻. The card appears in the library with a ▶ 试听 player and a · 未激活 badge until you synthesize with it.

Text to audio — paste text (≤10 000 characters), pick a voice, adjust speed / volume / pitch / emotion, synthesize. The mp3 is written to ~/.dsh/minimax-voice/output/ and the response says where.

Audio to text — pick a file, optionally name the language (zh, en, …), press 开始转写. The transcript lands in a read-only box with a copy button, and in out_path if you used the tool.

Agent tools

Tool	What it does
minimax_voice_clone	Clone from audio_path; returns voice_id plus a local preview path
minimax_voice_list	List saved voices with voice_id, name, model, preview path
minimax_tts	Synthesize text with voice_id (defaults to the newest clone)
minimax_transcribe	Transcribe audio_path to plain text, optionally write out_path

// minimax_voice_clone
{ "audio_path": "~/samples/narrator.m4a", "name": "悬疑旁白", "need_noise_reduction": true }

// minimax_tts
{ "text": "第一章，雨夜。", "voice_id": "mvabc123", "model": "speech-2.8-hd",
  "speed": 1.1, "emotion": "calm", "out_path": "~/out/ch1.mp3" }

// minimax_transcribe
{ "audio_path": "~/out/ch1.mp3", "language": "zh", "out_path": "~/out/ch1.txt" }

HTTP API

All routes live under /mmvoice/api on the Harness web server (http://127.0.0.1:58500 by default), are unauthenticated, and are only meaningfully reachable from localhost.

Method	Route	Body / query	Returns
GET	/status	—	{ configured, keySource, dataDir, defaultModel, voiceCount, previewMode }
GET	/voices	—	{ voices: Voice[] }
GET	/voices/:id/audio	—	preview mp3
DELETE	/voices/:id	—	{ removed: true } (local only)
POST	/key	{ key }	{ configured: true }
POST	/clone	multipart or raw audio, ?name=&model=&previewText=&noiseReduction=1	Voice
POST	/tts	{ text, voiceId, model, speed, vol, pitch, emotion }	{ path, durationMs, characters, traceId }
POST	/transcribe	raw audio, ?filename=&language=	{ text, durationMs, traceId }
GET	/output/:file	—	generated audio
GET	/remote	—	voices reported by /v1/get_voice

curl -X POST --data-binary @sample.mp3 -H "content-type: audio/mpeg" \
  "http://127.0.0.1:58500/mmvoice/api/transcribe?filename=sample.mp3&language=zh"

Cordis service

Other plugins can inject the service instead of re-implementing the API:

export const inject = ['minimaxVoice']

export function apply(ctx) {
  ctx.tools.register({
    name: 'audiobook_render',
    async execute(args) {
      const { path } = await ctx.minimaxVoice.synthesize({
        voiceId: args.voiceId,
        text: args.paragraph,
        model: 'speech-2.8-hd',
      })
      return { path }
    },
  })
}

Method	Signature
status()	{ configured, keySource, dataDir, defaultModel, voiceCount, previewMode }
list() / get(voiceId) / remove(voiceId)	voice registry access
audio(voiceId)	{ bytes, path, mime } for the local preview
clone(request)	{ audioPath | bytes, filename, name, model, previewText, needNoiseReduction, promptAudioPath, promptText, signal, exec }
synthesize(request)	{ text, voiceId, outPath, model, speed, vol, pitch, emotion, format, signal, exec }
transcribe(request)	{ audioPath | bytes, filename, language, outPath, signal, exec }
setApiKey(key)	persists the key to secret.json
remoteVoices()	account-side voice list from MiniMax

Voice:

{ voiceId, name, model, createdAt, durationMs, previewPath, sampleName, activated }

activated === false means the voice exists locally and can be previewed, but has never been synthesized: MiniMax has not yet charged the ¥9.9 cloning fee.

Billing

Item	When it is charged
Voice cloning ¥9.9 / voice	On the first synthesis with the new voice_id, not at clone time
Preview audio	TTS characters only (the default sentence is a fraction of a cent)
Text to speech	Per TTS character, depending on model (speech-2.8-hd / -turbo, …)
Speech to text	Per ASR usage, billed separately from TTS

Check MiniMax pricing for current numbers. The panel shows characters for every synthesis so you can track spend.

Data and privacy

~/.dsh/minimax-voice/
├── voices.json          voice registry (source of truth)
├── secret.json          API key, plain text
├── audio/<id>.mp3       clone previews
└── output/*.mp3         synthesized audio

• secret.json is plain text. Protect the file, or supply the key through MINIMAX_API_KEY instead.
• Audio and text are sent to MiniMax to do the work; nothing else leaves the machine.
• remove() deletes the local registry entry and preview only. Delete on MiniMax's side from the console if you also want the voice gone there.

Local API security

/mmvoice/api can read the API key, write audio files and trigger billed calls, so every request passes the harness trust fence: connection.requestRejection() checks the Host/Origin pair (DNS rebinding) and requires an authenticated browser session or API token. A request that fails is answered with 401 or 403. If the connection service is unavailable, only loopback, same-origin requests are accepted. Another website therefore cannot reach these routes to spend your MiniMax balance or swap the key.

Troubleshooting

Symptom	Fix
configured: false	No key found; save one in the panel or export MINIMAX_API_KEY
Panel missing after install	Toggle the plugin off/on in Plugins, or restart the app
mmvoice/api/status returns 404	Plugin is not loaded in this profile — check dsh plugin --profile <name> list
mmvoice/api/* returns 401	No authenticated session for that request; open the panel from the harness UI (the routes are not open to other origins)
New voice shows 未激活	Expected. It is never been synthesized yet, so the fee is not paid
Clone fails with 400	Sample too short/long/quiet, or the account has no voice-cloning permission
Transcription returns empty text	Language hint conflicts with the audio, or the file is mostly music

Development

git clone https://github.com/Sunshine334419520/dsh-minimax-voice.git
dsh plugin --profile desktop add "$PWD"

Host code is plain ESM (index.js), the panel is a Cordis client bundle (client.js). No build step, no runtime dependencies. After editing, reload the plugin: disable and re-enable it in Plugins, then refresh the page for panel changes.

License

MIT © 2026 Sunshine334419520