Community Models

The models in supported model names are maintained or selected by the onnx-asr project. The wider community also publishes compatible models on Hugging Face, including fine-tunes for additional languages and alternative quantizations.

Browse all models tagged onnx-asr

Use the full Hugging Face repository ID to load a community model:

import onnx_asr

model = onnx_asr.load_model("xezpeleta/parakeet-tdt-0.6b-v3-basque-onnx-asr")
print(model.recognize("test.wav"))

Important

This is a curated snapshot inspected on July 16, 2026, not an exhaustive registry or compatibility guarantee. The entries below passed metadata and file-layout inspection only: they were not downloaded, executed, benchmarked, or checked for transcription quality. Community models are maintained and supported by their publishers.

Language fine-tunes

These repositories convert distinct upstream fine-tunes rather than copying an existing onnx-asr export.

Repository	Owner	Language	Architecture	Upstream model	Precision	License	Verification
xezpeleta/parakeet-tdt-0.6b-v3-basque-onnx-asr	xezpeleta	Basque	NeMo Conformer TDT	itzune/parakeet-tdt-0.6b-v3-basque	FP32, INT8 encoder	CC-BY-4.0	Metadata inspected
alefiury/parakeet-tdt-0.6b-v3-ptBR-TAGARELA-onnx	alefiury	Brazilian Portuguese	NeMo Conformer TDT	alexandreacff/parakeet-tdt-0.6b-v3-ptBR-plus	FP32	CC-BY-4.0	Metadata inspected
Buttermilk03/parakeet-primeline-onnx	Buttermilk03	German	NeMo Conformer TDT	primeline/parakeet-primeline	FP32, INT8	CC-BY-4.0	Metadata inspected
AlinClaudiu/SpeD-ParakeetRo-110M-onnx	AlinClaudiu	Romanian	NeMo Conformer CTC	gabrielpirlo/SpeD_ParakeetRo_110M_TDT-CTC	FP32	Apache-2.0	Metadata inspected
AigizK/GigaAM-Bashkir-CV25-ONNX	AigizK	Bashkir	GigaAM Multilingual CTC	AigizK/GigaAM-Bashkir-CV25	FP32, INT8	MIT	Metadata inspected

Orukeet multilingual TDT

Orukeet is a Parakeet TDT v3 fine-tune for 25 European languages, published under CC BY-SA 4.0. Its combined INT8 ONNX export works with the existing nemo-conformer-tdt loader. The Hugging Face repository contains several formats, so select the combined ONNX subdirectory with subfolder when calling load_model.

The example pins the release and keeps downloads in the Hugging Face cache:

import onnx_asr

model = onnx_asr.load_model(
    "oruk/orukeet",
    subfolder="onnx/combined-v0.1.0-int8",
    revision="eac739d754bb171287930e6e63386f5b88f8179e",
    quantization="int8",
    providers=["CPUExecutionProvider"],
)
print(model.recognize("test.wav"))

Use a 16 kHz mono WAV file. Only model files are downloaded; recognition runs locally. See the model repository for weight license and attribution files.

This is offline transcription, without translation or streaming partial results. The example has been checked with repeated speech recognition in English, German, Spanish and French, silence, and a fully cached run with networking disabled. These are integration checks, not an accuracy or latency comparison.

OpenVoiceOS model families

OpenVoiceOS publishes a large coordinated set of conversions. The table uses one representative repository for each family; follow its owner link to find other languages and model sizes.

Family and representative	Languages	Architecture	Upstream publisher	Precision in representative	License	Verification
AI4Bharat IndicConformer	Indic languages	NeMo Conformer CTC	AI4Bharat	FP32, INT8	MIT	Metadata inspected
IISc Vaani FastConformer	Multilingual and individual Indic languages	NeMo Conformer TDT	ARTPARK-IISc	FP32, INT8	MIT	Metadata inspected
NVIDIA monolingual Conformer	Multiple languages	NeMo Conformer CTC/RNN-T	NVIDIA	FP32, INT8	CC-BY-4.0	Metadata inspected
Localized Parakeet	Polish, Estonian, Dutch, Slovenian, Portuguese, and others	NeMo Conformer TDT	Community fine-tunes	FP32	CC-BY-4.0	Metadata inspected
Wav2Vec2	Multiple languages	Wav2Vec2 CTC	Multiple publishers	FP32	Apache-2.0	Metadata inspected

Licenses can differ between repositories in a family. Always check the selected model card and its upstream model before redistribution or commercial use.

Optimized variants

Some community repositories retain an existing model but provide a materially different ONNX optimization. These are listed separately from language fine-tunes.

Repository	Owner	Language	Architecture	Upstream model	Precision or optimization	License	Verification
Olicorne/parakeet-tdt-0.6b-v3-smoothquant-onnx	Olicorne	Multilingual	NeMo Conformer TDT	istupakov/parakeet-tdt-0.6b-v3-onnx	FP32, FP16, SmoothQuant INT8	CC-BY-4.0	Metadata inspected
gvij/parakeet-tdt-0.6b-v3-onnx-static-qdq-pc	gvij	Multilingual	NeMo Conformer TDT	nvidia/parakeet-tdt-0.6b-v3	Static QDQ per-channel INT8	CC-BY-4.0	Metadata inspected

Curation criteria

A repository is included when it is public and ungated, declares a model type supported by onnx-asr, contains a valid config.json, and provides the expected ONNX graphs and tokenizer or vocabulary files. It must also identify its purpose, upstream lineage, and license.

A different fine-tuned upstream model, language, architecture, or documented quantization method is considered meaningful. Unchanged forks, mirrors, re-uploads, incomplete repositories, and repositories that differ only by owner are omitted. Large coordinated conversion sets are represented as families to keep this page maintainable.

Metadata inspection does not validate ONNX graph integrity or recognition quality. If a model fails to load, report the problem to its publisher and see the troubleshooting guide.