LongCat AudioDiT — Pinokio Launcher

Diffusion-based text-to-speech with zero-shot voice cloning, built on LongCat-AudioDiT from Meituan.

What it does

• Voice cloning: Upload a reference audio clip, auto-transcribe it, type new dialogue, and generate speech in that voice
• Zero-shot TTS: Generate speech without a reference clip
• Two models: 1B (lighter, ~6-8GB VRAM) and 3.5B (higher quality, ~16-20GB VRAM)
• Auto-transcription: Reference audio is transcribed via Groq Whisper API (free) or local faster-whisper

How to use

1. Click Install in Pinokio to set up dependencies
2. Click Start to launch the web UI
3. Upload a reference audio clip (optional — for voice cloning)
4. Click Auto-Transcribe Reference to get the transcript
5. Type your target text in the right panel
6. Adjust model, guidance method, steps, and strength as needed
7. Click Generate to produce the audio

Settings

Groq API (optional, recommended)

Set GROQ_API_KEY as an environment variable for fast cloud-based transcription. Without it, the app falls back to local faster-whisper (slower but works offline).

API

Python

import requests

# Voice cloning
with open("reference.wav", "rb") as f:
    ref_audio = f.read()

response = requests.post("http://localhost:7860/api/generate", json={
    "text": "The new line I want spoken",
    "ref_audio": "<base64 encoded audio>",
    "ref_transcript": "What the reference clip says",
    "model_id": "meituan-longcat/LongCat-AudioDiT-1B",
    "guidance_method": "apg",
    "steps": 16,
    "strength": 4.0,
    "seed": 1024
})

JavaScript

const response = await fetch("http://localhost:7860/api/generate", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    text: "The new line I want spoken",
    model_id: "meituan-longcat/LongCat-AudioDiT-1B",
    guidance_method: "apg",
    steps: 16,
    strength: 4.0,
    seed: 1024
  })
});

cURL

curl -X POST http://localhost:7860/api/generate \
  -H "Content-Type: application/json" \
  -d '{"text": "Hello world", "model_id": "meituan-longcat/LongCat-AudioDiT-1B", "guidance_method": "apg", "steps": 16, "strength": 4.0, "seed": 1024}'

Limitations

• Maximum audio duration: 30 seconds per generation
• Languages: English and Mandarin Chinese
• Voice cloning requires both reference audio AND its transcript
• Models are loaded in F32 — the 3.5B model needs significant VRAM

Credits

• LongCat-AudioDiT by Meituan (MIT License)
• Groq for Whisper transcription API
• faster-whisper for local transcription fallback