Buttermilk03/parakeet-primeline-onnx

Parakeet Primeline ONNX

This repository contains an ONNX export of primeline/parakeet-primeline for use with onnx-asr and Scriber.

The source model is a German-focused Parakeet/FastConformer-TDT ASR model based on NVIDIA Parakeet TDT 0.6B v3. The original model card reports 2.95% average WER across its German benchmarks and is licensed under CC-BY-4.0.

Files

• config.json
• vocab.txt
• encoder-model.onnx and decoder_joint-model.onnx for full precision fp32
• ONNX external data weight files used by the fp32 encoder export
• encoder-model-int8.onnx and decoder_joint-model-int8.onnx for the compact CPU-oriented int8 export
• scriber-quantizations.json

Quantizations

Quantization	Files	Notes
int8	encoder-model-int8.onnx, decoder_joint-model-int8.onnx	CPU-load validated with onnx-asr; preferred for smaller downloads and normal local use.
fp32	encoder-model.onnx, decoder_joint-model.onnx, external data files	Full precision export. Load from a complete repository snapshot because the encoder uses many external data files.

Scriber usage

Scriber can load this model through its Local ONNX provider without installing NeMo/Torch on end-user machines. Scriber selects from the validated int8 and fp32 variants and uses the complete local Hugging Face snapshot for fp32.

Attribution

Source model: primeline/parakeet-primeline, licensed CC-BY-4.0. This repository only repackages the model as ONNX for local inference.