README.md

---
license: mit
tags:
  - coreml
  - audio
  - speech-enhancement
  - noise-suppression
---

# DeepFilterNet3 for Core ML (streaming, 100 frames, FP16)

A Core ML conversion of [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet) by Hendrik Schröter, used for noise reduction in Clarity Sound Studio, a free, open-source macOS app.

Only the network is converted. The DSP around it (STFT, ERB and complex features, the mask, the deep filter and the inverse STFT) runs in the app. `dfn3_dsp.json` records that contract: window, ERB band widths, normalisation constants and the coefficient layout. The same JSON is in the model's `user_defined_metadata["dsp"]`.

| | |
| --- | --- |
| File | `DeepFilterNet3-Stream-T100-fp16.mlpackage` (ML Program, FP16, macOS 14+) |
| Upstream | DeepFilterNet3 weights at commit `d375b2d` (`dfn3-epoch120-best`) |
| Conversion | `convert.py --frames 100 --precision fp16 --gru unrolled`: stateful in chunks of 100 frames (10 ms hop at 48 kHz), the GRU and convolution state passed in and out as tensors |
| Inputs | `feat_erb`, `feat_spec` and six state tensors (zero at the start) |
| Outputs | `mask`, `coefs` and the next state |

The weights are unchanged apart from the conversion to FP16.

## Licence

DeepFilterNet is dual-licensed under MIT or Apache-2.0. This conversion is distributed under the MIT licence; see `LICENSE`.

Copyright (c) 2021 Hendrik Schröter.

> H. Schröter, T. Rosenkranz, A. N. Escalante-B. and A. Maier, "DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement", Interspeech 2023.