Skip to content

Installation

The package can be installed from PyPI.

Quick Install

# CPU only / CoreML for Apple Silicon (simplest install, minimal native deps)
pip install onnx-asr[cpu,hub]

# With NVIDIA GPU support (requires installed CUDA/cuDNN/TensorRT)
pip install onnx-asr[gpu,hub]

# Using uv
uv pip install onnx-asr[cpu,hub]

Requirements

ONNX Runtime Packages

onnx-asr requires an ONNX Runtime package. There are several options depending on your OS and hardware:

Package Providers Notes
onnxruntime CPUExecutionProvider, CoreMLExecutionProvider Default, works on all platforms
onnxruntime-gpu CPUExecutionProvider, CUDAExecutionProvider, TensorrtExecutionProvider For NVIDIA GPUs (requires NVIDIA libs)
onnxruntime-directml, onnxruntime-windowsml (new) CPUExecutionProvider, DmlExecutionProvider DirectML - only for Windows but no additional deps
onnxruntime-webgpu (beta) CPUExecutionProvider, WebGpuExecutionProvider WebGPU - cross-platform and no additional deps

Additional ONNX Runtime packages (not tested):

  • onnxruntime-trt-rtx - NVIDIA TensorRT for RTX
  • onnxruntime-qnn - Qualcomm Snapdragon
  • onnxruntime-openvino - Intel OpenVINO
  • onnxruntime-rocm - AMD GPUs (legacy)
  • onnxruntime-migraphx - AMD GPUs (new)
  • onnxruntime-cann - Huawei Ascend NPU

Note

Only onnxruntime and onnxruntime-gpu have predefined extras ([cpu] and [gpu]). Other packages can be installed manually.

Supported Platforms

The supported platforms are primarily determined by available ONNX Runtime wheels.

OS / CPU Python 3.10 Python 3.11 Python 3.12 Python 3.13 Python 3.14 Python 3.13t Python 3.14t
Linux x86_64 ✅ ✅ ✅ ✅ ✅ ✅ ✅
Linux Arm64 ✅ ✅ ✅ ✅ ✅ ✅ ✅
Windows x86_64 ✅ ✅ ✅ ✅ ✅ ❌ ❌
Windows Arm64 ❌ ✅ ✅ ✅ ✅ ❌ ❌
macOS Arm64 (Apple Silicon) ✅ ✅ ✅ ✅ ✅ ❌ ❌
macOS x86_64 ✅ ✅ ✅ ✅ ❌ ❌ ❌

Minimum Dependency Versions

Package Minimum Version
numpy 1.22.4
onnxruntime 1.18.1 (or any ONNX Runtime package)
huggingface-hub 0.30.2 (optional, for model downloading)
typing-extensions 4.6.0 (Python < 3.11 only)

Note

Older versions of ONNX Runtime packages may work but are not tested. The minimum version listed is the one used in CI testing.

Install from PyPI

  1. With CPU onnxruntime and huggingface-hub:

    pip install onnx-asr[cpu,hub]

  2. With onnxruntime for NVIDIA GPUs and huggingface-hub:

    pip install onnx-asr[gpu,hub]

    Warning

    First, you need to install the required version of CUDA/cuDNN for CUDAExecutionProvider and required TensorRT for TensorrtExecutionProvider (optional).

    You can also install onnxruntime CUDA/cuDNN dependencies and TensorRT via pip:

    pip install onnxruntime-gpu[cuda,cudnn] tensorrt-cu12-libs

  3. With onnxruntime for WinML and huggingface-hub:

    pip install onnx-asr[hub] onnxruntime-windowsml

  4. Without onnxruntime and huggingface-hub (if you already have some version of onnxruntime installed and prefer to download the models yourself):

    pip install onnx-asr

Install from source

To install the latest version of onnx-asr from sources, use pip (or uv pip):

pip install git+https://github.com/istupakov/onnx-asr