Parakeet TDT 0.6B v3

Parakeet TDT 0.6B v3 is a 600-million-parameter multilingual automatic speech recognition (ASR) model developed by NVIDIA and released in August 2025. It extends its predecessor parakeet-tdt-0.6b-v2 from English to 25 European languages, automatically detecting the spoken language without additional prompting.

The architecture combines a FastConformer encoder with a Token-and-Duration Transducer (TDT) decoder. The model adds automatic punctuation and capitalization, produces word- and segment-level timestamps, and supports long-form audio — up to 24 minutes with full attention (on an A100 80GB GPU) or up to 3 hours with local attention. Input is 16 kHz mono audio (.wav, .flac); output is text.

The model was trained on a combination of the Granary dataset's ASR subset (about 660,000 hours of pseudo-labeled data) and the in-house NeMo ASR Set 3.0 (about 10,000 hours of human-transcribed data). It is part of the Parakeet family within the NVIDIA NeMo toolkit and is released under the CC BY 4.0 license for commercial and non-commercial use.