@surfmate.team/digital-human-minimax

MiniMax adapters for the digital-human suite — they implement the ports from @surfmate.team/digital-human-ports:

• VoiceSynthesisPort — cloud TTS (t2a_v2), via a key-holding proxy (secure) or browser-direct (dev, key exposed).
• VoiceCloningPort — sample → cloned voiceId (upload → voice_clone → use), plus list / delete.

npm i @surfmate.team/digital-human-minimax

TTS

import { createMinimaxSynthAdapter } from '@surfmate.team/digital-human-minimax'

// Secure: a proxy holds the key server-side and returns mp3 bytes.
const synth = createMinimaxSynthAdapter({ endpoint: import.meta.env.VITE_TTS_PROXY_URL })

const bytes = await synth.synthesize({
  text: '你好',
  voiceId: 'minimax:male-qn-jingying',
  modelId: 'speech-02-hd', // optional
  emotion: 'sad',          // optional
})

// Dev only (key in the browser):
// createMinimaxSynthAdapter({ apiKey, groupId, host })

createMinimaxSynthAdapter(config)

config is one of:

| Mode | Shape | Notes | | --- | --- | --- | | proxy (secure) | { endpoint } | POSTs { text, voiceId, model, emotion }; the proxy adds the key and returns mp3 bytes. The key never reaches the browser. | | direct (dev) | { apiKey, groupId, host? } | Calls MiniMax t2a_v2 from the caller. ⚠️ key ships in the bundle — dev only. host defaults to api.minimax.io. |

Returns a VoiceSynthesisPort: synthesize(input) · playUrl(url) · playBytes(bytes) · stop().

synthesize(input) — VoiceSynthesisInput

| Field | Type | Required | Notes | | --- | --- | --- | --- | | text | string | ✅ | What to speak. | | voiceId | string | ✅ | minimax:<id> — a preset (minimax:male-qn-jingying) or a cloned id (minimax:surf…). The minimax: prefix is stripped before the API call. | | modelId | string | — | Only speech-02-hd or speech-02-turbo are honored. Anything else (or omitted) → speech-02-turbo. | | emotion | string | — | Only angry / sad / fearful / disgusted / surprised / neutral are honored. happy → neutral (it shifts the timbre); any other / omitted → no emotion sent. |

Returns Promise<ArrayBuffer> (mp3 bytes) — cache them and replay via playBytes.

The modelId / emotion whitelists are deliberate: a stale or foreign value (e.g. a mate-imported expressive model) is coerced instead of 502-ing the call.

Voice cloning (browser-direct, dev)

import { createMinimaxVoiceCloning } from '@surfmate.team/digital-human-minimax'

const cloning = createMinimaxVoiceCloning({ apiKey, groupId })
const { voiceId } = await cloning.cloneFromSample(file)  // 'minimax:surf…'
const mine = await cloning.listClones()                  // hides foreign/malformed ids

from-env helpers: createMinimaxSynth(env) / createMinimaxCloning(env) pick proxy-vs-direct from loose env config.

License

MIT