████████╗████████╗███████╗
╚══██╔══╝╚══██╔══╝██╔════╝
   ██║      ██║   ███████╗
   ██║      ██║   ╚════██║
   ██║      ██║   ███████║
   ╚═╝      ╚═╝   ╚══════╝

WinLab TTS

Text to speech in cloned voices for lab members, with the OpenAI API: change the base URL and the OpenAI SDK speaks in your own voice.

nextjs indextts openai-api postgres docker mcp

   

Members sign in with their lab account. Speech runs on a lab GPU with IndexTTS-2, so no text or voice leaves the lab.

> "Say 'the meeting moves to 2 pm' in CC's voice"
  ⚡ speak { "input": "會議改到下午兩點", "voice": "cctseng" }
✓ { "url": "https://tts.winlab.tw/a/…/….mp3", "seconds": 2.1 }

What it does

• Speaks text in a voice through POST /v1/audio/speech, the page, or MCP.
• Clones a member's own voice from a consent recording and a sample of up to 60 seconds, recorded in the browser or sent to /v1/audio/voice_consents and /v1/audio/voices. Each member has one voice, named after their lab account; recording again replaces it.
• Checks consent: Whisper must hear the fixed consent sentence, and a speaker model must find the sample and the consent recording to be the same person.
• Keeps a voice private until its owner shares it with every member; the owner can delete it, and its recordings go with it.
• Lets an admin add a member's voice with a note of how the speaker agreed, such as Prof. Tseng's (cctseng), owned by that member.

Use

Make a key at https://tts.winlab.tw/keys, then:

curl https://tts.winlab.tw/v1/audio/speech \
  -H "Authorization: Bearer $TTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "indextts-2", "input": "大家好", "voice": "cctseng"}' \
  -o speech.mp3

from openai import OpenAI

client = OpenAI(base_url="https://tts.winlab.tw/v1", api_key=TTS_API_KEY)
client.audio.speech.create(model="indextts-2", voice="cctseng", input="大家好").write_to_file("speech.mp3")

• voice: the owner's lab account name (cctseng), a voice id (voice_…), or {"id": "voice_…"}.
• response_format: mp3 (default), opus, aac, flac, wav or pcm (24 kHz, 16-bit mono).
• speed: 0.25 to 4.
• model is accepted and ignored: there is one model.
• input: up to 1,000 characters. A sentence takes a few seconds.

Your own voice, as OpenAI's custom voices:

# 1. Read the sentence for your language (GET /v1/audio/voice_consents lists them).
curl https://tts.winlab.tw/v1/audio/voice_consents -H "Authorization: Bearer $TTS_API_KEY" \
  -F language=zh -F recording=@consent.wav
# 2. A sample of the same speaker, 60 seconds or less.
curl https://tts.winlab.tw/v1/audio/voices -H "Authorization: Bearer $TTS_API_KEY" \
  -F audio_sample=@sample.wav -F consent=cons_… -F shared=true

A name is ignored: the voice takes your name. shared=true lets every member speak in it. GET /v1/audio/voices lists the voices you can use; DELETE /v1/audio/voices/<id> deletes yours.

MCP

https://tts.winlab.tw/mcp, with OAuth sign-in through the lab account. Tools: whoami, list_voices, speak (returns a link that works for 24 hours), list_speeches, delete_speech, share_voice, delete_voice, list_voice_consents, my_usage, list_api_keys, revoke_api_key.

How it works

browser / SDK / MCP ──▶ web (Carrel, tts.winlab.tw) ──▶ Postgres: keys, voices, jobs
                                     ▲
                                     │ long-poll /api/worker/next, post results
                                     │
                         worker (GPU box behind NAT): IndexTTS-2, Whisper, ECAPA

The worker only makes outbound HTTPS calls, so the GPU box needs no open port. A request waits on its job row until the worker posts the audio.

Deploy

The web app is one image with Postgres beside it:

cp .env.example .env   # fill it in
docker compose up -d

The worker runs on any machine with an NVIDIA GPU (16 GB is plenty; IndexTTS-2 peaks at 7.4 GiB):

cd worker
hf download IndexTeam/IndexTTS-2 --local-dir models/IndexTTS-2
printf 'TTS_URL=https://tts.example.com\nWORKER_TOKEN=…\n' > .env
docker compose up -d --build

Configuration

See .env.example. WORKER_TOKEN must match on both sides. TTS_ADMINS lists who may add a member's voice from a consent note.

Limitations

• One worker serves one job at a time; a busy queue answers 429 past 3 waiting requests per member.
• No streaming: the audio comes back whole.
• Audio links and results are kept for 24 hours.

Develop

bun install
bun run typecheck && bun run lint && bun run actions:check && bun run build

License

MIT