VoiceVibe/parakeet-ultra-nounk-onnx

parakeet-ultra-nounk (ONNX)

ONNX export of moondream/parakeet-ultra, a post-trained version of nvidia/parakeet-tdt-0.6b-v3, in the file layout used by onnx-asr. It is the offline speech-recognition model of the VoiceVibe desktop app.

Usage

import onnx_asr

model = onnx_asr.load_model("nemo-parakeet-tdt-0.6b-v3", path="path/to/this/repo")
print(model.recognize("speech.wav"))

The model type string stays nemo-parakeet-tdt-0.6b-v3: the architecture, tokenizer and graph are the same as v3's. Only the weights differ.

Changes from the original

• Converted the Moondream weights (Hugging Face safetensors, fp16) to a NeMo checkpoint in fp32, reusing NVIDIA v3's preprocessor, tokenizer and config.
• Suppressed the <unk> token: its output bias in the joint network is lowered by 10,000, so the model never emits <unk> (shown as "⁇"). The original emits it where its training text had characters the tokenizer can't spell, such as brackets, dashes, quotes and &.
• Exported to ONNX with NeMo's export (opset 17), following the recipe of istupakov/parakeet-tdt-0.6b-v3-onnx. The same recipe applied to v3 reproduces those files byte for byte.

Accuracy

Word error rate (%) through onnx-asr 0.7.0 in fp32, on FLEURS test clips, scored against the raw transcriptions with Whisper's normalizers. Absolute numbers are higher than leaderboard figures, which use prepared references; compare the two columns with each other.

FLEURS	parakeet-tdt-0.6b-v3 (ONNX)	this model
English (647 clips)	6.70	6.06
Spanish (300)	3.78	3.18
Portuguese (300)	5.77	4.96
German (300)	7.31	6.73

Speed and memory are the same as v3's.

License and attribution

CC BY 4.0. This model is adapted from:

1. "Parakeet Ultra" (moondream/parakeet-ultra, revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b) by Moondream, CC BY 4.0, which is itself a post-trained version of
2. "parakeet-tdt-0.6b-v3" (nvidia/parakeet-tdt-0.6b-v3) by NVIDIA Corporation, CC BY 4.0.

The material was modified as described above. No endorsement by Moondream or NVIDIA is implied. Provided "as is", without warranties, as stated in the license.