HE-NAC 300k 🎧
The 300k checkpoint for HE-NAC: High-Efficiency Neural Audio Coding Via Hierarchical and Residual Multi-Band Quantization. HE-NAC is a 32 kHz neural audio codec with core, mid-band, and high-band paths trained in sequence. Its encoder and residual vector quantizers build on DAC.
- Code and training recipes: darius522/henac on GitHub
- Paper: IEEE Transactions on Audio, Speech and Language Processing
- Project page: minjekim.com/research-projects/he-nac
- Checkpoint license: MIT (see LICENSE)
Checkpoint
| File | Contents | SHA-256 |
|---|---|---|
henac-300k.pth |
Generator weights and embedded model configuration | 978c7e9950b787b7a3673c198678c711cd5fab2fe766f4e41516125fc35a305e |
This is a self-contained inference checkpoint made from the original 300k high-band weights and their saved configuration. It does not include the discriminator, optimizer, or training corpus. The recommended operating point is 16 core, 4 mid, and 2 high codebooks.
Use it 🚀
From a checkout of the code repository:
python -m pip install -e .
python -m pip install -U huggingface_hub
hf download darius522/henac henac-300k.pth --local-dir checkpoints
henac-infer --checkpoint checkpoints/henac-300k.pth \
--input path/to/audio.wav --output outputs --codebooks 16 4 2
The CLI accepts WAV or FLAC input, resamples to 32 kHz, and writes reconstructed PCM WAVs. It averages multichannel input to mono for coding and copies the reconstruction back to the input channel count. See the code repository for staged training and evaluation commands.
Validation and scope
The consolidated implementation strictly loads this checkpoint. On 39 supplied
MUSHRA inputs, its [16, 4, 2] reconstructions were byte identical to
those regenerated by the original entropy_ctrl_hb implementation on the
checked A100 runtime. Against the supplied gold WAVs, the lowest exact PCM
fraction was 98.870% and the lowest signal-to-difference ratio was 47.96 dB;
the legacy implementation showed the same differences. See the
parity record
for hashes and test conditions.
The current code reconstructs waveforms but does not produce a compressed
bitstream. Its entropy.csv values are empirical Shannon entropy estimates,
not actual Huffman-coded file sizes. The paper reports bitrates measured after
Huffman coding. The original training corpus is not redistributed here, so a
full 300k-step training trajectory has not been reproduced.
Citation 📄
@ARTICLE{11718947,
author={Pétermann, Darius and Lim, Wootaek and Beack, Seungkwon and Kim, Minje},
journal={IEEE Transactions on Audio, Speech and Language Processing},
title={HE-NAC: High-Efficiency Neural Audio Coding Via Hierarchical and Residual Multi-Band Quantization},
year={2026},
volume={},
number={},
pages={1-15},
keywords={Modeling;Codes;Bit rate;Speech;Encoding;Audio coding;High frequency;Distance measurement;Frequency;Decoding;Neural audio coding;high-efficiency;multiband;residual},
doi={10.1109/TASLPRO.2026.3739174}}