Purevo Studio uses CEVA DPDFNet-8 48 kHz HR.
Official implementation: https://github.com/ceva-ip/DPDFNet
Official weights: https://huggingface.co/Ceva-IP/DPDFNet
Code and weights: Apache License 2.0.

Core ML conversion by Kyle Howells:
https://huggingface.co/iky1e/DPDFNet8-48kHz-HR-CoreML
Pinned revision: c92bc188144d3167ee0d345fe9c4e80b9a5d66eb
The model's LICENSE and attribution README ship in Resources/Studio/.

The native runtime implements the published spectrum/state interface, Vorbis
window and 2,400-sample compensation. The conversion's reference runtime is
https://github.com/kylehowells/DeepFilterNet-mlx (MIT OR Apache-2.0).
DeepFilterNet-mlx: Copyright 2026 Kyle Howells.

The independently written room-reflection filter and voice/output processing
are by Echo App Studio. Room reduction is delayed linear-prediction DSP, not
another neural model. It cannot recover a clipped or missing word.
