mohammedali2501/ssondo

S-SONDO

Self-Supervised Knowledge Distillation for General Audio Foundation Models

ICASSP 2026

     

Authors: Mohammed Ali El Adlouni*, Aurian Quelennec*, Pierre Chouteau, Geoffroy Peeters, Slim Essid

LTCI, Telecom Paris, Institut Polytechnique de Paris

---

S-SONDO distills large audio foundation models into lightweight students that are up to 61x smaller while retaining up to 96% of teacher performance — using only output embeddings, no logits or layer-level alignment required.

Overview of the S-SONDO framework. Student embeddings are mapped and aligned with teacher embeddings in the teacher's latent space through self-supervised knowledge distillation.

Install

pip install ssondo

Quick Start

from ssondo import get_ssondo

# Load model (auto-downloads and caches)
model = get_ssondo()

# Extract embeddings from raw audio (mono, 32kHz)
embeddings = model(audio)  # (batch, n_segments, 960)

That's it. No preprocessing, no config files, no manual downloads.

Pretrained Classifiers

7 ready-to-use classifiers, trained via linear probing on standard audio benchmarks:

model = get_ssondo(head="esc50")
logits = model(audio)  # (batch, 50) — environmental sound classification

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

from ssondo import list_heads
print(list_heads())  # see all available heads

Custom Classification Heads

Attach your own head for any downstream task:

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

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

# Freeze backbone for linear probing
model.freeze_backbone()
model.train()

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

Results

Downstream evaluation across 7 audio tasks (4 music + 3 environmental sound). Students retain up to 96.4% of teacher performance while being up to 61x smaller.

	w/ BDS	wo/ BDS
MATPAC++ → MobileNetV3	73.0	72.7
MATPAC++ → DyMN	72.6	72.9
MATPAC++ → ERes2Net	70.8	44.8
M2D → MobileNetV3	69.2	69.4
M2D → DyMN	68.7	69.2
M2D → ERes2Net	69.2	68.7

Impact of Balanced Data Sampling (BDS) on downstream performance. BDS is critical for ERes2Net, improving from 44.8 to 70.8 average accuracy.

Available Checkpoints

from ssondo import list_models
print(list_models())  # see all available backbones

Model	Teacher	Student	Params	Emb. Size
matpac-mobilenetv3	MATPAC++	MobileNetV3	2.9M	960
matpac-dymn	MATPAC++	DyMN	8.7M	960
matpac-eres2net	MATPAC++	ERes2Net	1.4M	10240
m2d-mobilenetv3	M2D	MobileNetV3	2.9M	960
m2d-dymn	M2D	DyMN	8.7M	960
m2d-eres2net	M2D	ERes2Net	1.4M	10240

API Reference

from ssondo import get_ssondo

# Embeddings only (default)
model = get_ssondo()
embeddings = model(audio)                    # (batch, n_segments, 960)
emb = model.get_embeddings(audio)            # (batch, 960) — mean-pooled

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

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

# Finetuning
model.freeze_backbone()                      # linear probing
model.unfreeze_backbone()                    # full finetuning

# GPU
model = get_ssondo(device="cuda")

# Local checkpoint
model = get_ssondo("path/to/checkpoint.ckpt")

# Properties
model.embedding_dim                          # 960
model.backbone                               # raw nn.Module

Input Requirements

Parameter	Value
Format	Raw mono waveform
Sample rate	32,000 Hz
Window	10s segments (automatic)
Spectrogram	128 mel bands, 50–16000 Hz

Training Code

Full 4-step training pipeline to reproduce all results:

git clone https://github.com/MedAliAdlouni/ssondo_temp
cd ssondo_temp/training_ssondo
./setup.sh          # install deps, download models
./run_pipeline.sh   # end-to-end demo

See the training README for details.

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