TypeWhisper/typewhisper-win model-parakeet-ultra-int8-v2

Parakeet Ultra 0.6B model files (int8, v2)

on GitHub

Model files the NVIDIA Parakeet plugin downloads for Parakeet Ultra 0.6B. This is not an app or plugin release.

moondream/parakeet-ultra, Moondream's retrained nvidia/parakeet-tdt-0.6b-v3, exported to the sherpa-onnx transducer layout with eng/export_parakeet_sherpa.py (torch 2.8.0, nemo_toolkit 3.0.0, onnx 1.18.0, onnxruntime 1.22.1). The export started from the same NeMo checkpoint as model-parakeet-ultra-int8-v1, parakeet-tdt-0.6b-v3-ultra.nemo from Olicorne/parakeet-tdt-0.6b-v3-ultra-onnx at revision b27f93c2238aa35e138996c005a62f1e8430f801.

The quantization differs from v1 in the encoder only: its MatMuls are quantized to int8 and its convolutions stay in float. The decoder, joiner and tokens.txt are byte-identical to v1. On FLEURS (25 languages × 10 clips) this lowers the word error rate from 14.9 % to 11.5 % and transcribes faster on x64; the download grows from 671 MB to 898 MB.

License: CC-BY-4.0. Model by Moondream, derived from NVIDIA's parakeet-tdt-0.6b-v3.

SHA-256:

• encoder.int8.onnx 04f2c974fddac63b39620d4c834d3a99c4270aec7d16725e9a4081a70b105bc9
• decoder.int8.onnx 1fab98fe6c12aded87d2da66272cc9e148d0a0044ce3850a12fe56302ec4a922
• joiner.int8.onnx 8a71aaccfdba3d451775507a889afd2257d2e754463fd81095a8c8f290f924c2
• tokens.txt d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d

(#781)