Egorgij21/Audio-SAE-HuBERT-large

Audio-SAE — HuBERT-large

BatchTop-K Sparse Autoencoders trained on every transformer layer of facebook/hubert-large-ll60k, from the paper AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders (EACL 2026).

Each SAE decomposes the residual stream at one encoder layer into a sparse, largely interpretable dictionary of features.

• Code: https://github.com/audiosae/audiosae_demo
• Paper: https://arxiv.org/abs/2602.05027
• Collection: https://huggingface.co/collections/Egorgij21/audio-sae

Specs

Backbone	Activation dim	Dict size	Expansion	k	Layers
HuBERT-large	1024	8192	8×	50	24

One SAE per encoder layer (layer_1 … layer_24). Layer indices are 1-based and correspond to the output of the n-th transformer block.

Layout

layer_1/
  ae.pt          # BatchTopKSAE state_dict
  config.json    # training config (activation_dim, dict_size, k, …)
layer_2/
…
layer_24/

Each ae.pt contains encoder.{weight,bias}, decoder.weight, b_dec, k.

Loading

import torch
from huggingface_hub import hf_hub_download
from audio_sae import BatchTopKSAE
from audio_sae.models.hubert import MyHubert

device = "cuda" if torch.cuda.is_available() else "cpu"
layer = 12

# 1. HuBERT-large encoder, tapped after `layer`
hubert = MyHubert("facebook/hubert-large-ll60k", sae_after_layer=layer).to(device).eval()

# 2. Matching SAE
ckpt = hf_hub_download(
    repo_id="Egorgij21/Audio-SAE-hubert-large",
    filename=f"layer_{layer}/ae.pt",
)
sae = BatchTopKSAE.from_pretrained(ckpt, device=device)

# 3. Run on audio
import librosa
wav, _ = librosa.load("example.wav", sr=16000, mono=True)
wav = torch.from_numpy(wav).unsqueeze(0).to(device)

with torch.no_grad():
    acts = hubert(wav)                                # (1, T, 1024)
    features = sae.encode(acts, use_threshold=True)   # (1, T, 8192), sparse

See the GitHub repo for a full inference and interpretability walkthrough.

Training

• Architecture: BatchTop-K SAE, 8× expansion
• Optimizer: Adam, lr 2e-4, 200 000 steps, decay from step 160 000
• Loss: L2 reconstruction with batch-wide top-k (k=50)
• Data: ~2.8 k hours of mixed audio — speech (LibriSpeech, LibriHeavy, IEMOCAP, ESD, Expresso, CREMA-D, MELD), music (MTG-Jamendo) and environmental sounds (MUSAN, DEMAND, WHAM, FSD50K, VocalSound, Nonspeech7k, ESC-50, VGGSound)
• Seed: 21

See the paper for full training details and evaluation metrics.

Citation

@inproceedings{aparin2026audiosae,
  title     = {AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders},
  author    = {Aparin, Georgii and Sadekova, Tasnima and Rukhovich, Alexey and Yermekova, Assel and Kushnareva, Laida and Popov, Vadim and Kuznetsov, Kristian and Piontkovskaya, Irina},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year      = {2026},
  address   = {Rabat, Morocco},
}

License

MIT