NVIDIA parakeet-tdt-0.6b-v3
Release in the NVIDIA Parakeet family · version parakeet-tdt-0.6b-v3
Maintained by NVIDIA (NeMo)16
parakeet-tdt-0.6b-v3 is a 600-million-parameter multilingual speech recognition model with a FastConformer encoder and a TDT decoder, which predicts tokens and their durations jointly. It extends parakeet-tdt-0.6b-v2 from English to 25 European languages, detects the spoken language automatically, and outputs punctuated, capitalized text with word- and segment-level timestamps. NVIDIA released it on Hugging Face on August 14, 2025.13
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- License: CC-BY-4.0 legal code (external site: creativecommons.org)
- Paper: Canary-1B-v2 & Parakeet-TDT-0.6B-v3 technical report (arXiv 2509.14128) (external site: arxiv.org)
- Repository: NeMo Speech repository (external site: github.com)
- Dataset hub: Granary dataset (external site: huggingface.co)
Availability and license
Overall availability
Downloadable from Hugging Face without an access gate; the card states the model is ready for commercial and non-commercial use.21
Availability is separate from permission: read the license before using or redistributing.
Creative Commons Attribution 4.0 International (external site: creativecommons.org)124
Apache License 2.0 (NeMo Speech toolkit) (external site: raw.githubusercontent.com)5
The card's governing terms put the model under CC-BY-4.0, which grants a worldwide, royalty-free license to reproduce, share, and adapt the material and requires attribution and license notices when sharing it; the card also states the model is ready for commercial and non-commercial use. The NeMo toolkit used to run and train it is a separate Apache 2.0 repository.145
Component reuse rights
- weights
- Terms or license combinations require review
- code
- Reviewed qualifying license recorded — check scope and conditions
- data
- Unknown — no complete fact-level rights review
- documentation
- Unknown — no complete fact-level rights review
No complete system-rights review is recorded for this release.
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | The ungated repository holds safetensors and .nemo checkpoints plus a q8_0 GGUF file.2 |
| Inference codeIs code for running the model published? | Public | The card documents inference with NVIDIA NeMo, the NeMo-Speech.cpp C++ runtime, and Hugging Face Transformers, including timestamps, long-form, and streaming examples.15 |
| Training codeIs the code used to train the model published? | Public | The card links the NeMo transducer training script and FastConformer hybrid TDT-CTC configuration it was trained with; both are in the Apache 2.0 NeMo Speech repository.1785 |
| Training-data informationDoes the information cover provenance, scope, acquisition, selection, labeling, processing, and where data or alternatives can be obtained? Access alone does not establish completeness. | Partial | The card describes about 660,000 hours of pseudo-labeled Granary data (from YTC, MOSEL, and YODAS) and about 10,000 hours of human-transcribed NeMo ASR Set 3.0 data, listing the component corpora; the Granary card and paper describe the labeling pipeline. This review confirms disclosure, not completeness across every provenance, labeling, and processing requirement.1910 |
| Training-data accessCan the training data be obtained? This is independent of information completeness and reuse rights; original unshareable data need not be downloadable. | Partial | Granary is published as CC-BY-4.0 transcript manifests that point to separately downloaded source audio; NeMo ASR Set 3.0 is described as an in-house set, and access terms for its component corpora were not reviewed.91 |
| Complete training pipelineIs the complete base-training and preprocessing pipeline published, including configuration? Fine-tuning code or an inference SDK alone is insufficient. | Unknown | Not assessed. |
| Legacy data assessment (v0.1)Historical assessment combining download access and disclosure. Preserved for traceability; excluded from the v0.2 tier calculation. See the new separate assessments above. | Unknown | Not assessed. |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The card gives the initialization (an NVIDIA CTC multilingual checkpoint pretrained on Granary), step counts and GPU counts for both stages, temperature-based data balancing, and the tokenizer setup; the recipe for the initial checkpoint is not described there.1 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The card publishes word error rates on FLEURS, MLS, CoVoST, and Open ASR Leaderboard datasets; evaluation scripts were not verified.1 |
What it is useful for
The card lists speech-to-text uses such as conversational AI, voice assistants, transcription services, subtitle generation, and voice analytics, and documents long-form transcription with a local-attention mode (up to three hours of audio).1
Run and use notes
- The card lists NeMo 2.4 as the runtime, Linux as the supported operating system, and NVIDIA Volta, Ampere, Hopper, and Blackwell GPUs as compatible, and shows a local NeMo-Speech.cpp workflow using the q8_0-quantized GGUF file.1
Organization context
Provenance and derivatives
Trained by NVIDIA from an NVIDIA CTC multilingual checkpoint pretrained on Granary. Granary was built by the NVIDIA NeMo team with CMU and FBK teams; its card says transcripts were pseudo-labeled with Whisper-large-v3 and that punctuation and capitalization were restored with Qwen-2.5-7B, both third-party models. No third-party model weights are documented as part of this checkpoint.19
Other releases in the NVIDIA Parakeet family
- NVIDIA parakeet-unified-en-0.6bModel-disclosure tier (USASI rubric v0.2): Open-weight
What this catalog does not know
- Complete training pipeline: unknown.
- Legacy data assessment (v0.1): unknown.
Have a primary source? How to report a correction.
U.S. eligibility
Eligible · basis: U.S. headquarters
Published by NVIDIA in its Hugging Face organization, trained with NVIDIA's NeMo toolkit per the card, and announced in NVIDIA's NeMo Speech repository. NVIDIA's principal executive offices are in Santa Clara, California, per its Form 10-Q for the quarter ended July 26, 2026. Third-party models used to label training data are recorded under provenance.1611
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.