Paraformer-large (ONNX)

Non-autoregressive ASR with ONNX runtime — optimized for deployment without PyTorch dependency.

This repository contains the ONNX-exported version of Paraformer-large for efficient CPU/GPU inference via ONNX runtime.

Quick Start (PyTorch)

from funasr import AutoModel

model = AutoModel(
    model="funasr/paraformer-zh",
    hub="hf",
    vad_model="funasr/fsmn-vad",
    punc_model="funasr/ct-punc",
    device="cuda",
)
result = model.generate(input="audio.wav")
print(result[0]["text"])

ONNX Runtime Usage

from funasr_onnx import Paraformer

model = Paraformer("funasr/Paraformer-large", quantize=True)
result = model(audio_in="audio.wav")
print(result)

Features

• ONNX format for cross-platform deployment
• Quantized model for faster CPU inference
• No PyTorch dependency needed at runtime
• Same accuracy as PyTorch version

Model Details

Property	Value
Architecture	Paraformer (Non-autoregressive)
Parameters	220M
Format	ONNX
Languages	Chinese, English
Training Data	60,000+ hours

Links

• GitHub: FunASR
• PyTorch version: funasr/paraformer-zh
• Paper: Paraformer: Fast and Accurate Parallel Transformer