Upstream only ships the .nemo bundle, and transformers' own convert_nemo_to_hf.py script doesn't handle v2's 1024-token vocab (it hardcodes a labels.index("<pad>") assumption that only holds for v3). This repo is the conversion output with that one-line fallback applied — see live-llm-bench/scripts/convert-parakeet-v2-nemo-to-hf.py for the exact transform. Weights are identical to upstream; only the packaging is changed.

Usage

python

from transformers import AutoModelForTDT, AutoProcessor

model_id = "ai-and-i-project/parakeet-tdt-0.6b-v2-hf"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTDT.from_pretrained(model_id, dtype="auto", device_map="cuda")

License

Inherits cc-by-4.0 from nvidia/parakeet-tdt-0.6b-v2.