Parakeet Ultra — MLX fp16
moondream/parakeet-ultra (revision
73175eb7), converted to the MLX layout of
beshkenadze/parakeet-tdt-0.6b-v3-mlx-fp16
for mlx-audio-swift. Parakeet Ultra is Moondream's
post-trained nvidia/parakeet-tdt-0.6b-v3:
same architecture and vocabulary, better accuracy.
Conversion
convert_parakeet_ultra.py renames the transformers tensors to the MLX names, transposes the
convolutions to MLX order, sums each LSTM layer's two biases, drops the VAD head, and casts to
fp16. config.json and vocab.txt are the v3 MLX files. The script refuses a result unless
every v3 tensor is filled with one of the same shape, and unless each is close to its v3
counterpart (cosine similarity at least 0.9; measured minimum 0.954, median 1.000).
Accuracy
WER through the same MLX pipeline (Silero VAD, mlx-audio-swift), against the v3 MLX fp16 weights this replaces:
| Set | v3 | Ultra |
|---|---|---|
| FLEURS, 13 languages × 5 clips | 6.60% | 5.71% |
| FLEURS Russian, 61 clips | 4.97% | 4.31% |
| A 5-minute Russian meeting, one track | 27.0% | 24.2% |
License
CC-BY-4.0, as the upstream weights. Parakeet Ultra by Moondream; Parakeet TDT 0.6B v3 by NVIDIA.