Pretrained audio CNNs that actually beat Google's baseline

What it does

This is the research code behind PANNs — a family of CNNs trained on Google’s AudioSet (527 sound classes, ~5,000 hours). You get pretrained weights for audio tagging and frame-wise sound event detection, plus scripts to fine-tune on your own task. The best model, Wavegram-Logmel-CNN, hits 0.439 mAP on AudioSet; the simpler Cnn14 gets 0.431. Both outperform the Google baseline of 0.317.

The interesting bit

The authors trained directly on raw audio waveforms rather than relying on Google’s provided embeddings, which was the common pattern at the time. They also ship a separate panns_inference package for users who just want pretrained models without the full training pipeline — a rare case of research code acknowledging that most people only want the weights.

Key highlights

• Pretrained models available on Zenodo; inference works out of the box with pytorch/inference.py
• Supports both clip-level tagging and frame-level sound event detection (DecisionLevelMax, Avg, Att variants)
• Fine-tuning template included; demonstrated on GTZAN music classification
• Training from scratch takes 3–7 days on a single V100 (works on 12 GB GPUs with reduced batch size)
• Full reproducibility package: dataset download scripts, HDF5 packing for I/O speedup, and plotting utilities for all paper figures

Caveats

• The AudioSet download is ~1.1 TB and hosted on YouTube, so expect missing files and geographic variability
• Baidu Cloud link provided for the authors’ exact downloaded copy; no direct HTTP mirror
• README is thorough but reads like a lab notebook — you’ll need to dig for the high-level picture

Verdict

Worth a look if you need off-the-shelf sound classification or event detection and don’t want to train on AudioSet yourself. Skip if you’re after real-time streaming or edge deployment — these are research models, not optimized inference engines.