PANNs
Large-scale pre-trained audio neural networks; CNN14 and Wavegram-Logmel-CNN on AudioSet.
About
PANNs asked a simple question: what if audio pattern recognition had an ImageNet moment? Kong and colleagues trained a family of CNNs (CNN6/10/14, ResNets, MobileNets, and the waveform-plus-spectrogram Wavegram-Logmel-CNN) on the full AudioSet corpus and released the code and checkpoints. The best system hit 0.439 mAP on AudioSet tagging, a clear step over the previous best of 0.392 at the time, and the models transferred well to six downstream tasks including music classification.
The reference implementation lives at qiuqiangkong/audioset_tagging_cnn under MIT, checkpoints are hosted on Zenodo, and there is a friendly panns-inference PyPI wrapper plus a training script template for fine-tuning. PANNs became one of the standard transfer-learning backbones in audio research, which means plenty of community knowledge and third-party ports. The headline 0.439 mAP is a provider-reported 2020 result; the ecosystem has since moved partly to transformers, but CNN14 remains a strong, cheap baseline.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: 0.439 mAP on AudioSet tagging (TASLP 2020), SOTA at publication; transferred to six audio tasks with SOTA on several. Numbers are provider-claimed.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified, metrics-v1). Evidence: peer-reviewed TASLP 2020 paper (arXiv 1912.10211); MIT-licensed repo qiuqiangkong/audioset_tagging_cnn (~1787 stars, recently active); Zenodo-hosted checkpoints; panns-inference PyPI wrapper; heavy citation and reuse. Strengths: open everything, strong transfer results, mature community tooling. Docked for: (1) headline mAP is a provider-reported 2020 figure; (2) transformer models have since taken some of its benchmarks. License: MIT (code); weights freely downloadable from Zenodo.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · TASLP 2020 (arXiv 1912.10211), Kong et al., University of Surrey: CNN14/ResNet/Wavegram-Logmel-CNN on full AudioSet; 0.439 mAP SOTA at publication
- · Repo qiuqiangkong/audioset_tagging_cnn (~1787 stars, recently active); Zenodo checkpoints; panns-inference PyPI wrapper; fine-tune template
- · Headline mAP is a provider-reported 2020 figure; transformers have since taken some of its benchmarks