Practical overview
When it fits
Use it when a Python workflow needs speech recognition through ONNX models and benefits from built-in preprocessing and decoding, a documented API or CLI, and model options for local or Hugging Face loading.
Who it is for
- Python developers adding speech-to-text to applications or server-side components.
- Developers building command-line transcription utilities.
- Teams selecting among supported ONNX speech-recognition model families.
Documented uses
- Transcribe speech from WAV files or NumPy audio arrays.
- Run batch recognition or process longer recordings using voice activity detection.
- Load supported models from Hugging Face or local directories, including quantized models.
- Retrieve token-level timestamps and log probabilities when needed.
Requirements & limits
The repository identifies the project as open source under the MIT license. The documented installation uses pip; the quick start uses the CPU and Hugging Face Hub extras.
- NumPy and ONNX Runtime are required; Hugging Face Hub is optional.
- The project lists Python 3.10–3.14 and NumPy 1.22–2.4+ support.
- Most models have a 20–30-second maximum audio length unless long-form recognition uses VAD.
- Supported devices and execution providers vary; the project lists Windows, Linux and macOS, x86 and Arm CPUs, and CUDA, TensorRT, CoreML, DirectML, ROCm and WebGPU.
- Community models are maintained by their publishers and may not be tested by the project.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install onnx-asr and transcribe a WAV file
Run speech recognition on a WAV file using the Parakeet TDT v3 model.
Before you start
- Python 3.10 or later
- A WAV file named test.wav in the working directory
- Internet access to retrieve the model from Hugging Face
Steps
Install the CPU and Hugging Face Hub extras with the command below.
pip install onnx-asr[cpu,hub]In Python, import
onnx_asrand load the model withmodel = onnx_asr.load_model("nemo-parakeet-tdt-0.6b-v3").Transcribe the file and print the result with
result = model.recognize("test.wav")followed byprint(result).
Expected outcome
The model returns a transcription result for the WAV file, which is printed.
Guide limitations
- Model retrieval may take a few minutes.
- Most models have a maximum audio length of 20–30 seconds; use VAD for longer audio.
- Direct WAV reading supports PCM_U8, PCM_16, PCM_24, and PCM_32. Other formats must first be converted or read into a NumPy array using a library.
Transcribe a WAV file with the onnx-asr command-line interface
Use the CLI to recognize speech in a WAV file with the Parakeet TDT v3 model.
Before you start
- onnx-asr installed
- A WAV file named test.wav in the working directory
Steps
Run
onnx-asr nemo-parakeet-tdt-0.6b-v3 test.wav.For CLI usage parameters, consult the CLI help with
onnx-asr -h.
Expected outcome
The CLI runs recognition on the specified WAV file.
Guide limitations
- Most models have a maximum audio length of 20–30 seconds; use VAD for longer audio.
- Direct WAV reading supports PCM_U8, PCM_16, PCM_24, and PCM_32. Other formats must first be converted or read into a NumPy array using a library.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 13 October 2026
- Original website title
- istupakov/onnx-asr · Official homepage title · retained in its source language