Parakeet-TDT-0.6b-v2 (fp16 ONNX)
An fp16 ONNX build of NVIDIA Parakeet-TDT-0.6b-v2, an English automatic speech recognition (ASR) model. This repository re-hosts the model in the ONNX format at half precision (fp16) so it can be run with ONNX-based runtimes. The model weights are unchanged in substance: this is a format and precision conversion, not a re-trained model.
Model description
- Original model: NVIDIA Parakeet-TDT-0.6b-v2.
- Architecture: FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. This is a transducer (RNN-T family) ASR model.
- Parameters: approximately 0.6 billion (600 million).
- Language: English.
- Task: speech to text, with punctuation and capitalisation.
- Audio input: 16 kHz mono audio.
- Format in this repository: ONNX, fp16 (half precision), split into the standard transducer parts (encoder, decoder, joiner) plus the tokeniser.
Why this variant exists
The other ONNX conversions of this model that are readily available are int8-quantised. This repository provides an fp16 build instead, which is needed to run on hardware and runtimes that require fp16 precision, such as the Intel NPU via OpenVINO.
Intended use
This model is intended for English speech-to-text transcription: for example, dictation, transcription of recordings, subtitle generation, and voice interfaces. It expects 16 kHz mono audio and produces text with punctuation and capitalisation. It is an English-only model and is not intended for other languages or for speaker identification.
Files
| File | Description |
|---|
encoder.fp16.onnx | The FastConformer acoustic encoder, in fp16 ONNX. This is the largest file (roughly 1.2 GB). |
decoder.fp16.onnx | The transducer prediction network (decoder), in fp16 ONNX. |
joiner.fp16.onnx | The transducer joint network (joiner) that combines encoder and decoder outputs, in fp16 ONNX. |
tokens.txt | The token table mapping model output indices to text tokens. |
tokenizer.model | The SentencePiece tokeniser model used to encode and decode text. |
tokenizer.vocab | The SentencePiece vocabulary listing. |
vocab.txt | A plain-text vocabulary listing. |
The tokeniser files (tokens.txt, tokenizer.model, tokenizer.vocab, vocab.txt) travel with the model and form part of it. They are covered by the same CC-BY-4.0 licence as the model. They are model artifacts and are separate from the Apache-2.0 SentencePiece software library, which is not included here.
How this was produced
The original NVIDIA .nemo model was converted to fp16 ONNX by the sherpa-onnx / k2-fsa (Next-gen Kaldi) project. The converted artifact is distributed by that project as a release asset, and the files in this repository are re-hosted from that conversion.
Provenance and attribution
- Original model: NVIDIA Parakeet-TDT-0.6b-v2, created and published by NVIDIA. See the source model card at https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2.
- ONNX conversion: performed by the sherpa-onnx / k2-fsa (Next-gen Kaldi) project. Their conversion toolkit is licensed under Apache-2.0. The conversion toolkit's licence applies to the tooling, not to the model weights.
- Change made: the model was converted from the original NVIDIA
.nemo model to fp16 ONNX. No re-training was performed. This notice is provided to satisfy the CC-BY-4.0 requirement to indicate that a change was made.
Licence
This model is distributed under the Creative Commons Attribution 4.0 International licence (CC-BY-4.0), the same licence as the original NVIDIA model. See
https://creativecommons.org/licenses/by/4.0/ for the human-readable summary, and the
LICENSE file in this repository for the full legal text.
Under CC-BY-4.0 you are free to share and adapt this model, including for commercial purposes, provided you give appropriate credit to NVIDIA, link back to the licence, and indicate if changes were made. Credit and the change notice are given in the sections above.
