lightware-dev/parakeet-tdt-0.6b-v3-bf16

parakeet-tdt-0.6b-v3 (bf16)

A bfloat16 cast of NVIDIA's nvidia/parakeet-tdt-0.6b-v3, packaged as a single .nemo archive for direct GPU loading.

The weights are identical to the upstream model apart from the fp32 → bf16 conversion. Restoring from this checkpoint loads straight onto a CUDA GPU without ever materialising an fp32 copy, which is faster to start (13 s vs ~22 s) and roughly halves peak load-time VRAM (1.3 GB). It is intended for bf16-capable CUDA GPUs; on hardware without bf16 support, use the upstream fp32 checkpoint.

This repository redistributes the model under its original CC BY 4.0 license. See the upstream model card for architecture, training data, supported languages, evaluation, and intended use.

Usage (NeMo)

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("lightware-dev/parakeet-tdt-0.6b-v3-bf16",
                       "parakeet-tdt-0.6b-v3-bf16.nemo")
model = nemo_asr.models.ASRModel.restore_from(ckpt, map_location="cuda")
model.eval()
print(model.transcribe(["audio.wav"]))

Provenance

• Base model: nvidia/parakeet-tdt-0.6b-v3
• Change: fp32 weights cast to bf16, re-saved as .nemo
• License: CC BY 4.0 (inherited from the base model)