funasr-onnx-automodel 0.1.0

AutoModel wrapper for funasr-onnx with unified VAD→ASR→Punc pipeline

funasr-onnx-automodel

A unified AutoModel interface for funasr-onnx that automatically orchestrates the VAD → ASR → Punc speech recognition pipeline with a single generate() call.

Why this project?

funasr-onnx provides excellent ONNX-based speech recognition models, but using them requires manually loading and chaining multiple models (VAD, ASR, Punc) with different APIs and data formats. This package provides a single unified interface — just like HuggingFace's AutoModel — so you can run the full pipeline in 3 lines of code.

Features

• One interface for all models — AutoModel automatically detects model type (Paraformer / SenseVoice)
• Full pipeline orchestration — VAD segmentation → ASR recognition → Punctuation restoration
• Flexible usage — Use full VAD+ASR+Punc pipeline, or ASR-only mode
• Quantized by default — Uses INT8 quantized models for fast CPU inference
• Bypasses funasr_onnx __init__.py — Avoids unnecessary heavy imports, faster startup

Installation

pip install funasr-onnx-automodel

Or with uv:

uv pip install funasr-onnx-automodel

Quick Start

Full Pipeline (VAD + ASR + Punc)

from funasr_onnx_automodel import AutoModel

model = AutoModel(
    model="iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-onnx",
    vad_model="iic/speech_fsmn_vad_zh-cn-16k-common-onnx",
    punc_model="iic/punc_ct-transformer_cn-en-common-vocab471067-large-onnx",
)

results = model.generate("audio.wav")
print(results[0]["text"])

ASR Only (no VAD, no Punc)

from funasr_onnx_automodel import AutoModel

model = AutoModel(
    model="iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-onnx",
)

results = model.generate("short_audio.wav")
print(results[0]["text"])

API Reference

AutoModel(model, vad_model=None, punc_model=None, quantize=True, device_id="-1", intra_op_num_threads=4, **kwargs)

Parameter	Type	Default	Description
model	str	required	Model ID or local path for ASR model. Model type (Paraformer/SenseVoice) is auto-detected from name.
vad_model	str | None	None	Model ID or local path for VAD model. If None, ASR runs without VAD segmentation.
punc_model	str | None	None	Model ID or local path for punctuation model. If None, no punctuation is added.
quantize	bool	True	Use INT8 quantized ONNX model (model_quant.onnx).
device_id	str | int	"-1"	Device ID. -1 for CPU inference.
intra_op_num_threads	int	4	Number of ONNX Runtime intra-op threads.

generate(input, **kwargs)

Parameter	Type	Description
input	str | list | np.ndarray	Audio file path, list of paths, or 16kHz float32 numpy array.
**kwargs		Passed through to the ASR model.

Returns: list[dict] — Each dict contains "text" (recognized text) and "timestamp" (word-level timestamps).

Model Download

Models are automatically downloaded from ModelScope on first use. Use the iic/ prefix for ModelScope model IDs. Pre-exported ONNX models have the -onnx suffix.

Recommended Models

Role	Model ID
VAD	iic/speech_fsmn_vad_zh-cn-16k-common-onnx
ASR	iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-onnx
Punc	iic/punc_ct-transformer_cn-en-common-vocab471067-large-onnx

Audio Format

Input audio should be 16kHz mono WAV. Use ffmpeg to convert:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

License

MIT

Acknowledgments

• FunASR by Alibaba DAMO Academy / ModelScope
• funasr-onnx for ONNX model runtime