onnx-asr 0.12.0

A lightweight Python package for Automatic Speech Recognition using ONNX models

ONNX ASR

onnx-asr is a Python package for Automatic Speech Recognition using ONNX models. It's a lightweight, fast, and easy-to-use pure Python package with minimal dependencies (no need for PyTorch, Transformers, or FFmpeg):

Key features of onnx-asr include:

• Supports many modern ASR models
• Runs on a wide range of devices, from small IoT/edge devices to servers with powerful GPUs (benchmarks)
• Works on Windows, Linux, and macOS on x86 and Arm CPUs, with support for CUDA, TensorRT, CoreML, DirectML, ROCm, and WebGPU
• Supports NumPy versions from 1.22 to 2.4+ and Python versions from 3.10 to 3.14
• Loads models from Hugging Face or local directories, including quantized versions
• Accepts WAV files or NumPy arrays, with built-in file reading and resampling
• Supports custom models (see the Conversion Guide for instructions)
• Supports batch processing
• Supports long-form recognition using VAD (Voice Activity Detection)
• Can return token-level timestamps and log probabilities
• Provides a fully typed and well-documented Python API
• Provides a simple command-line interface (CLI)

[!NOTE] Supports Parakeet v2 (En) / v3 (Multilingual), Canary v1/v2 (Multilingual), GigaAM v2/v3 (Ru), and GigaAM Multilingual models!

[!TIP] Try the onnx-asr demo on Hugging Face Spaces:

Quick Start

Install onnx-asr:

pip install onnx-asr[cpu,hub]

Load a model and recognize a WAV file:

import onnx_asr

# Load the Parakeet TDT v3 model from Hugging Face (may take a few minutes)
model = onnx_asr.load_model("nemo-parakeet-tdt-0.6b-v3")

# Recognize speech and print the result
result = model.recognize("test.wav")
print(result)

[!WARNING] The maximum audio length for most models is 20–30 seconds. For longer audio, use VAD.

For more examples, see the Usage Guide.

See the Installation Guide for detailed installation instructions.

Supported Model Architectures

The package supports the following modern ASR model architectures. See supported model names for a complete list and the model comparison for comparisons with the original implementations.

• NVIDIA NeMo Conformer/FastConformer/Parakeet/Canary (with CTC, RNN-T, TDT, and Transformer decoders)
• GigaChat GigaAM v2/v3/Multilingual (with CTC and RNN-T decoders, including E2E versions)
• Kaldi Icefall Zipformer (with a stateless RNN-T decoder), including Alpha Cephei Vosk 0.52+
• T-Tech T-one (with a CTC decoder; streaming is not yet supported)
• OpenAI Whisper

When these models are saved in ONNX format, typically only the encoder and decoder are included. Running them requires the corresponding preprocessing and decoding implementations. This package provides both for all supported models:

• Log-mel spectrogram preprocessors
• Greedy search decoding

Benchmarks

Inverse Real-Time Factor (RTFx) is the ratio of audio duration to processing time. An RTFx greater than 1 indicates faster-than-real-time processing; higher values indicate better performance.

Model	9800X3D CPU (RTFx)	Cortex A53 CPU (RTFx)	T4 CUDA (RTFx)	RTX 5070 Ti TensorRT (RTFx)
NeMo Parakeet v2/v3	36	1.0	57	320
NeMo Canary v2	8	N/A	21	36
GigaAM v3 CTC	59	1.6	84	1370
GigaAM v3 RNN-T	43	1.5	40	130

See the Benchmarks page for detailed performance benchmarks.

Troubleshooting and FAQ

See the Troubleshooting Guide for common issues and solutions.

For more help, check the GitHub Issues or open a new one.

License

MIT License