This is NVIDIA's parakeet-tdt-0.6b-v2 converted to the Hugging Face transformers format. NVIDIA distributes v2 as a .nemo checkpoint only, and to my knowledge no working HF-format conversion has been public before this one. The v2 release remains the strongest Parakeet for English transcription, and its newer multilingual sibling v3 trades some English accuracy for language coverage, so an HF-format v2 is worth having.

These weights are the starting point of a Core AI port that reaches 291x real time on a Mac and 209x on an iPhone at full precision, written up in A 10-hour audiobook, transcribed on an iPhone in 5 minutes.

The weights were converted from the official .nemo checkpoint using the transformers conversion script models/parakeet/convert_nemo_to_hf.py, with two fixes that the script needs for v2 (a <pad> crash and a silent blank-id off-by-one that corrupts transcripts). The fixes have been submitted upstream: https://github.com/huggingface/transformers/pull/47891.

Verification against the NeMo/PyTorch reference: a greedy TDT decode over this checkpoint matches the reference generate() output token for token on probe audio (82 of 82 tokens), and end-to-end transcripts were token-exact across a multi-hour audiobook validation corpus.

Key config facts: vocab_size=1025, blank_token_id=1024, durations=[0, 1, 2, 3, 4], joint activation ReLU, 128-mel front end with 0.97 preemphasis. The encoder is architecturally identical to v3's, and only the tokenizer and vocab differ between the two releases. Note that 1025 is the id-space size (max id plus one) rather than the token entry count, since <unk> reuses id 0.

The original model is by NVIDIA and is licensed CC-BY-4.0. This repository redistributes converted weights under the same license with attribution. Changes made: format conversion from .nemo to safetensors/transformers layout as described above, with no change to the weights themselves.