ncannings/parakeet-ultra-nemo

parakeet-ultra in NeMo format

moondream/parakeet-ultra (Moondream's post-trained nvidia/parakeet-tdt-0.6b-v3, 25 languages) converted to a NeMo .nemo checkpoint, so it runs in NVIDIA NeMo and in fastconformer-trt.

The weights are Moondream's, unchanged and in full precision. The conversion places every tensor into the v3 NeMo model and refuses unless every NeMo parameter is filled exactly once with a matching shape. Moondream's small voice-activity head is not part of the NeMo architecture and is left out, so use your own segmentation for long audio. Conversion script: engine/ultra_to_nemo.py.

Use with NeMo

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

path = hf_hub_download("ncannings/parakeet-ultra-nemo", "ultra.nemo")
model = nemo_asr.models.ASRModel.restore_from(path)
print(model.transcribe(["speech.wav"])[0].text)

Fast inference: fastconformer-trt

fastconformer-trt builds an FP8 TensorRT engine with custom CUDA plugins from this checkpoint. TensorRT engines are specific to the GPU and TensorRT version, so no engine is published here: build it on the machine that will run it (instructions in the repository).

GPU	Stock NeMo	fastconformer-trt	Speed-up
NVIDIA DGX Spark (GB10)	999x	4,903x	4.9x
NVIDIA GH200	3,873x	22,776x	5.9x
NVIDIA H100 SXM	4,693x	20,667x	4.4x
NVIDIA B200	5,743x	25,073x	4.4x

Real-time factor on LibriSpeech test-clean, timed from audio already in GPU memory to text. Accuracy matches stock NeMo on the same scorer: FLEURS mean over 25 languages 12.60% against 12.55%, full Earnings-22 10.07% against 10.03%. These are scored with the Whisper normalisers, so they differ from the Open ASR Leaderboard figures on Moondream's model card; compare only within one scorer. Full method and per-language results are in the repository.

Licence and attribution

CC-BY-4.0, as the original. Model by Moondream, based on NVIDIA parakeet-tdt-0.6b-v3. Converted by Nigel Cannings; the only change is the file format.