ssondo 0.1.1

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

S-SONDO

Lightweight audio embeddings from self-supervised knowledge distillation.

S-SONDO provides compact audio models (MobileNetV3, DyMN, ERes2Net) trained via knowledge distillation from large audio foundation models (MATPAC, M2D). Extract general-purpose audio embeddings with a single function call.

Paper: S-SONDO: Self-Supervised Knowledge Distillation for General Audio Foundation Models (ICASSP 2026)

Installation

pip install ssondo

Quick Start

import torchaudio
from ssondo import get_ssondo

# Load a pretrained model (auto-downloads from Hugging Face Hub)
model = get_ssondo("matpac-mobilenetv3")

# Load audio (mono, 32kHz)
x, sr = torchaudio.load("audio.wav")
x = x.mean(dim=0, keepdim=True)  # mono

# Extract embeddings
embeddings = model(x)  # (1, n_segments, 960)

Available Models

from ssondo import list_models

for name, description in list_models().items():
    print(f"{name}: {description}")

Model	Teacher	Student	Embedding Size
matpac-mobilenetv3	MATPAC++	MobileNetV3	960
matpac-dymn	MATPAC++	DyMN	960
matpac-eres2net	MATPAC++	ERes2Net	varies
m2d-mobilenetv3	M2D	MobileNetV3	960
m2d-dymn	M2D	DyMN	960
m2d-eres2net	M2D	ERes2Net	varies

Usage

Extract Embeddings

model = get_ssondo("matpac-mobilenetv3")
embeddings = model(audio)  # (batch, n_segments, emb_size)

Get Logits Too

model = get_ssondo("matpac-mobilenetv3", return_logits=True)
embeddings, logits = model(audio)

GPU Inference

model = get_ssondo("matpac-mobilenetv3", device="cuda")
embeddings = model(audio.cuda())

Load from Local Checkpoint

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

Finetuning with Frozen Backbone (Linear Probe)

import torch
from ssondo import get_ssondo

model = get_ssondo("matpac-mobilenetv3")
model.freeze_backbone()  # freeze all backbone params
model.train()

# Add a linear classifier for your task
head = torch.nn.Linear(model.embedding_dim, num_classes)

# Extract embeddings (backbone frozen, no grad)
emb = model.get_embeddings(audio)  # (batch, 960)
logits = head(emb)
loss = criterion(logits, labels)
loss.backward()  # only head parameters are updated

Full Finetuning

model = get_ssondo("matpac-mobilenetv3")
model.train()  # all parameters trainable by default

Useful Properties

model.embedding_dim   # 960 — size of backbone embeddings
model.backbone        # the raw backbone nn.Module (e.g., MobileNetV3)

Input Requirements

• Mono audio (single channel)
• Sample rate: 32,000 Hz
• Audio is internally sliced into 10-second segments and converted to 128-band log-mel spectrograms

How It Works

get_ssondo() auto-detects everything from the checkpoint: student backbone, preprocessing parameters, and classification head. No manual configuration needed.

When you pass a model name (e.g., "matpac-mobilenetv3"), the checkpoint is automatically downloaded from Hugging Face Hub and cached locally.

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