Mega-ASR — Pinokio Launcher

One-click installer for Mega-ASR, a foundation speech recognition model built for robust transcription in challenging real-world audio environments.

What it does

Mega-ASR transcribes audio accurately even in noisy, far-field, echoing, or reverberant conditions. It was trained on 2.6M samples across 54 acoustic scenarios using a two-stage approach (supervised fine-tuning + reinforcement learning).

How to use

1. Click Install in Pinokio — this clones the repo, installs dependencies, and downloads the model (~8GB VRAM recommended)
2. Click Start — launches the Gradio web UI
3. Upload an audio file or record from your microphone
4. Get your transcription

Requirements

• GPU with 8GB+ VRAM (NVIDIA recommended, Apple Silicon supported)
• ~10GB disk space for model weights + dependencies

API

Python

from transformers import AutoModelForCausalLM, AutoProcessor
import librosa

model = AutoModelForCausalLM.from_pretrained("zhifeixie/Mega-ASR", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("zhifeixie/Mega-ASR", trust_remote_code=True)

audio, sr = librosa.load("audio.wav", sr=16000)
inputs = processor(audio=audio, sampling_rate=sr, return_tensors="pt")
outputs = model.generate(**inputs)
text = processor.batch_decode(outputs, skip_special_tokens=True)[0]
print(text)

CLI

cd app
python infer.py --audio /path/to/audio.wav

cURL (when web UI is running)

curl -X POST http://localhost:7860/api/predict \
  -H "Content-Type: application/json" \
  -d '{"data": ["audio.wav"]}'

Credits

• Model & Code: xzf-thu/Mega-ASR
• Paper: arXiv
• HuggingFace: zhifeixie/Mega-ASR