onnx-asr

What it is and what it does

onnx-asr is a Python library for speech-to-text inference using ONNX-format models. It wraps modern ASR architectures (NeMo Conformer/Parakeet/Canary, GigaAM, Kaldi Icefall Zipformer, T-Tech T-one, and OpenAI Whisper) with preprocessing and decoding logic, letting you load a model and transcribe audio in a few lines of code. The package is designed to be lightweight—it requires only numpy and typing-extensions, avoiding heavy dependencies like PyTorch or Transformers—and runs on diverse hardware from IoT devices to GPU servers.

You provide either a WAV file or NumPy array, and the package handles resampling, log-mel spectrogram computation, and greedy-search decoding. It supports batch processing, Voice Activity Detection for long-form audio, token-level timestamps, and log probabilities. Models load from Hugging Face or local directories, including quantized versions. The library is actively maintained, fully typed, and includes both a Python API and a command-line interface.

Use it for:

• Transcribe audio files offline without cloud dependencies or heavy ML frameworks in production services.
• Build speech-to-text features on edge devices or IoT hardware with constrained resources.
• Process long-form audio recordings using Voice Activity Detection to split and recognize speech segments.
• Integrate multilingual ASR (Parakeet v3, Canary, GigaAM Multilingual) into applications requiring non-English transcription.
• Benchmark or compare ONNX-based ASR models across different hardware (CPU, CUDA, TensorRT, CoreML).
• Develop custom ASR pipelines by loading quantized or fine-tuned ONNX models from Hugging Face.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Automatic Speech Recognition using ONNX models with minimal dependencies, supporting multiple modern ASR architectures and running on CPUs, GPUs, and edge devices.

Yes. The package is actively maintained, has low install friction, carries a permissive MIT license, and solves a real problem—lightweight offline speech recognition without PyTorch or heavy dependencies. It supports modern models and diverse hardware. No known vulnerabilities. Install it if you need ASR inference in Python without framework overhead.

Install

onnx-asr on PyPI

pip install onnx-asr

uv add onnx-asr

poetry add onnx-asr

Installing onnx-asr

Before you install

Low friction: pure Python wheel with only numpy and typing-extensions as runtime dependencies. Active maintenance with a release 30 days ago and 361 repository stars. Supports Python 3.10 through 3.14.

License in practice

MIT license (permissive) allows commercial and private use with minimal restrictions.

Quickstart

pip install onnx-asr[cpu,hub]

import onnx_asr
model = onnx_asr.load_model("nemo-parakeet-tdt-0.6b-v3")
result = model.recognize("test.wav")
print(result)

Most models have a maximum audio length of 20–30 seconds; longer audio requires Voice Activity Detection (VAD).

Verify before relying

• Whether the package's claimed support for CUDA, TensorRT, CoreML, DirectML, ROCm, and WebGPU requires additional system libraries or environment setup beyond pip install.
• Performance characteristics on specific hardware (e.g., actual RTFx values on your target device) beyond the published benchmarks.
• Whether quantized model variants are automatically downloaded or require manual setup.

Package facts

Evidence: onnx_asr-0.12.0-py3-none-any.whl

Keywords: asr, onnx, speech-recognition, speech-to-text, stt

Tags

More Libraries packages