wasi-master/parakeet-ultra-executorch

Parakeet Ultra, ExecuTorch export

Moondream's Parakeet Ultra, a post-trained parakeet-tdt-0.6b-v3, exported for ExecuTorch v1.5.1 with the XNNPACK backend. Offline voice typing in WM Keyboard downloads it.

Path	What
xnnpack/int4/model.pte	8-bit dynamic activations, 4-bit weights in groups of 32 (8da4w)
xnnpack/int4/tokenizer.model	SentencePiece tokenizer, unchanged from parakeet-tdt-0.6b-v3

The program has the methods preprocessor, encoder, decoder_step and joint, the same as ExecuTorch's own parakeet-tdt-0.6b-v3 export, so its Parakeet runner runs this file unchanged.

How it was made

Moondream publish Ultra as Hugging Face transformers weights. Their tensors were renamed to NeMo's names with transformers' own NeMo conversion mapping, run in reverse, and loaded into parakeet-tdt-0.6b-v3's NeMo model. The mapping was first checked on v3 itself: every NeMo weight equals the tensor of that name in NVIDIA's transformers conversion. Ultra's voice-activity head has no place in the NeMo model and was left out. The model was then quantized and exported with ExecuTorch's examples/models/parakeet/export_parakeet_tdt.py at v1.5.1. manifest.json holds the pinned sources, file hashes and test transcripts. The script is tools/parakeet-ultra/export_ultra.py.

License

CC-BY-4.0, as the original. Parakeet Ultra is by Moondream, built on NVIDIA's parakeet-tdt-0.6b-v3. Changes: converted to NeMo's layout, quantized and exported to ExecuTorch.