ssondo 0.3.1

S-SONDO: Lightweight audio embeddings from self-supervised knowledge distillation

S-SONDO

Lightweight audio embeddings from self-supervised knowledge distillation.

Up to 61x smaller than teacher models, retaining up to 96% performance.

   

ICASSP 2026

---

Install

pip install ssondo

Quick Start

from ssondo import get_ssondo

model = get_ssondo()
embeddings = model(audio)  # (batch, n_segments, 960)

No preprocessing, no config files, no manual downloads. Pass raw mono audio at 32 kHz and get embeddings.

Pretrained Classifiers

7 ready-to-use classifiers trained on standard audio benchmarks:

model = get_ssondo(head="esc50")
logits = model(audio)  # (batch, 50)

Head	Task	Classes
esc50	Environmental sound	50
us8k	Urban sound	10
fsd50k	Sound events	200
gtzan	Music genre	10
openmic	Instrument recognition	20
nsynth	Instrument family	11
magna-tag-a-tune	Music auto-tagging	50

Custom Heads

# Linear
model = get_ssondo(head="linear", n_classes=10)

# MLP
model = get_ssondo(head="mlp", n_classes=10, hidden_sizes=[512, 256])

Finetuning

# Linear probing (frozen backbone)
model = get_ssondo(head="linear", n_classes=10)
model.freeze_backbone()
model.train()

logits = model(audio)
loss = criterion(logits, labels)
loss.backward()  # only head parameters update

# Full finetuning
model.unfreeze_backbone()

API at a Glance

from ssondo import get_ssondo, list_models, list_heads

model = get_ssondo()                          # load model
model = get_ssondo(head="esc50")              # pretrained classifier
model = get_ssondo(head="linear", n_classes=10)  # custom head
model = get_ssondo(device="cuda")             # GPU
model = get_ssondo("path/to/checkpoint.ckpt") # local checkpoint

embeddings = model(audio)                     # (batch, n_segments, 960)
emb = model.get_embeddings(audio)             # (batch, 960) mean-pooled
model.embedding_dim                           # 960
model.backbone                                # raw nn.Module

list_heads()                                  # available classifiers

Model

S-SONDO ships with matpac-mobilenetv3 — a MobileNetV3 (2.9M params) distilled from MATPAC++, achieving the best downstream performance across all 7 benchmarks (96.4% of teacher performance at 61x fewer parameters). Embeddings are 960-dimensional.

Input

• Mono audio, single channel
• Sample rate: 32,000 Hz
• Internally sliced into 10 s segments and converted to 128-band log-mel spectrograms

Links

• Paper: arXiv
• Models: Hugging Face Hub
• Code & Training: GitHub

Citation

@inproceedings{eladlouni2026ssondo,
  title={S-SONDO: Self-Supervised Knowledge Distillation for General Audio Foundation Models},
  author={El Adlouni, Mohammed Ali and Quelennec, Aurian and Chouteau, Pierre and Peeters, Geoffroy and Essid, Slim},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2026}
}

License

MIT