parakeet-tdt-0.6b-v3 — fp16 ONNX
fp16 conversion of istupakov/parakeet-tdt-0.6b-v3-onnx
(same graph layout: encoder-model + fused decoder_joint-model, .fp16 suffix), made with
onnxruntime's transformers.float16 converter (keep_io_types=True, path-based shape inference
for the >2GB encoder). I/O stays fp32, weights/compute fp16.
Verified vs the fp32 export: random-input parity cos ≈ 1.0, and END-TO-END transcripts are byte-identical on DirectML while decoding ~2× faster (66 s clip: 231 ms vs 459 ms warm, RTX 3080 Ti; encoder 92 ms vs 300 ms). Half the download/VRAM of fp32.
Made for WinSTT.
- Downloads last month
- 66
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for Masterx/parakeet-tdt-0.6b-v3-fp16-onnx
Base model
nvidia/parakeet-tdt-0.6b-v3