speech-enhancement-native
Copyright 2026 BigLinux <www.biglinux.com.br>

The code in this repository is an independent implementation of the inference
path of the models below. It is licensed MIT OR Apache-2.0, except the
gtcrn-aec crates, which are Apache-2.0 (see each Cargo.toml). The model weights
are derived from the upstream releases and keep their upstream licenses. This
project is not affiliated with or endorsed by the upstream authors.

1. DeepFilterNet3 and DeepFilterNet3-LL
   Hendrik Schröter, Tobias Rosenkranz, Alberto N. Escalante-B., Andreas Maier.
   https://github.com/Rikorose/DeepFilterNet
   License: MIT OR Apache-2.0 (deepfilternet3/LICENSE-DeepFilterNet-MIT).
   Files: deepfilternet3/dfn3-ladspa/dfn3_weights.bin,
          deepfilternet3/dfn3ll-ladspa/dfn3ll_weights.bin
   Changes: GRU projection matrices quantized to int8 with per-row scales
   (deepfilternet3/scripts/quantize_int8.py); other tensors kept in f32 and
   repacked into the layout the engines load.

2. DPDFNet-2 48 kHz HR
   Daniel Rika, Nino Sapir, Ido Gus (Ceva). https://github.com/ceva-ip/DPDFNet
   License: Apache-2.0 (dpdfnet/model/LICENSE).
   Source code commit 9bd9844a227bb6aa57e55588d8d0e961fcff1c46.
   Checkpoint dpdfnet2_48khz_hr.pth from https://huggingface.co/Ceva-IP/DPDFNet,
   revision dd6818d00f50c836fed43a6243ebe49116de5964,
   sha256 f4eb984dc80d66d6e417a8d6ad8855b4fd0b559452db1e68950f028234109d5c.
   Files: dpdfnet/model/dpdfnet2_48khz_hr-w8a16/
   Changes: batch norm folded into the convolutions, GRU matrices quantized to
   int8 with per-row scales and reordered to z,r,n gates
   (dpdfnet/tools/export_model.py), then repacked losslessly
   (dpdfnet/tools/pack_matrices.py).

3. GTCRN-AEC (LocalVQE)
   localai-org. https://github.com/localai-org/LocalVQE
   License: Apache-2.0 (gtcrn-aec/model/LICENSE).
   File: gtcrn-aec/model/localvqe-pi-aec-v1-49k-f32.gguf, unmodified, from
   https://huggingface.co/LocalAI-io/LocalVQE,
   sha256 b80b75b9038d0d28079a84afe7d4f0b8f6404f723af4633c05ed4f96fb30b7b8.
   The gtcrn-aec crates port the LocalVQE inference code (ggml/gtcrn.cpp and
   ggml/daf_frontend.cpp) to Rust.
   Architecture: GTCRN, Xiaobin Rong et al., ICASSP 2024,
   https://github.com/Xiaobin-Rong/gtcrn (MIT).

4. Silero VAD v6.2.3 (16 kHz model)
   Silero Team. https://github.com/snakers4/silero-vad
   License: MIT (silero-vad/model/LICENSE-Silero).
   Source: silero_vad/data/silero_vad_16k_op15.onnx from the silero-vad 6.2.3
   Python wheel, sha256
   7ed98ddbad84ccac4cd0aeb3099049280713df825c610a8ed34543318f1b2c49.
   File: silero-vad/model/silero_vad_16k.bin
   Changes: the ONNX initializers written as raw little-endian f32
   (silero-vad/tools/export_weights.py), values unchanged.

5. FireRedVAD (AED and Stream-VAD)
   Kaituo Xu, Wenpeng Li, Kai Huang, Kun Liu (Xiaohongshu).
   https://github.com/FireRedTeam/FireRedVAD
   License: Apache-2.0 (firered-vad/model/LICENSE-FireRedVAD).
   Checkpoints AED/model.pth.tar, Stream-VAD/model.pth.tar and their cmvn.ark
   from https://huggingface.co/FireRedTeam/FireRedVAD,
   revision 7990aaccc6b7aec1e527743bd30201f2c4a03b8c, sha256
   ad08a4e05b58ca328154e158d24cff57a2fe796ceabb63bb701544c3f7d4f7ad (AED),
   ec88a8ae8ac5f004cdbd20c1eac4a9e9d12067c3d51ac5b141afa1f921fb59c9 (Stream-VAD),
   c87f6f13edf0f0ec7535ddfc9cc3387d9268cb234b70182d566c5e2edf3ca473 (cmvn.ark).
   Files: firered-vad/model/firered_aed.bin, firered-vad/model/firered_stream_vad.bin
   Changes: the CMVN statistics turned into means and inverse deviations, linear
   weights transposed and memory filters reordered by tap, written as raw
   little-endian f32 (firered-vad/tools/export_weights.py), values unchanged.

Rust dependencies keep their own licenses; `cargo deny check licenses` lists
them.
