PANNs CNN14 (AudioSet) — Core ML

PANNs CNN14 (Cnn14_mAP=0.431, Kong et al. 2020, 527-class AudioSet tagging; Zenodo 3987831) for Core ML on Apple devices.

Converted from the same verified reference used by the schism-mlx MLX ports. Two variants per model:

File	Precision	Compute units	max logit diff
`Cnn14_fp16.mlpackage`	FLOAT16	ALL (ANE)	1.3e-2, top-5 identical on tested clips
`Cnn14_fp32.mlpackage`	FLOAT32	CPU+GPU	2.9e-6

Verified on-device-equivalently via coremltools on an M5 Max, against the reference implementation on real audio. fp16 is ANE-eligible and recommended for iPhone / iPad; fp32 is the tight-parity fallback.

Download

.mlpackage bundles must be materialized as real files — the Core ML compiler rejects the symlinks that a default snapshot_download creates in the Hugging Face cache:

from huggingface_hub import snapshot_download
path = snapshot_download("schism-audio/cnn14-audioset-coreml", local_dir="./cnn14-audioset-coreml")

(or hf download schism-audio/cnn14-audioset-coreml --local-dir ./cnn14-audioset-coreml). Swift hosts downloading files directly are unaffected.

I/O contract

input logmel: (1, 1001, 64) float32 — librosa-style log-mel (32 kHz, n_fft 1024, hop 320, 64 mels, 50–14000 Hz, 10*log10(clamp(mel, 1e-10))) of a 10 s window
output logits: (1, 527) float32 — apply sigmoid; multi-label
longer audio: 10 s windows, aggregate scores

DSP frontend (host-side)

The Core ML graph contains the network only. The host implements the audio frontend and must match schism_mlx.audio numerically — test_vectors_*.npz in this repo holds deterministic input/output pairs (float32; match within ~1e-4 relative to be interchangeable with what these models were verified against). No chunking layer is needed for full tracks — CNN14 accepts any input length natively (windowing only sets score granularity). A validated Swift implementation (Accelerate; modules SchismDSP and SchismPipeline) is available at schism-audio/schism-dsp, tested against these exact vectors.

License

CC-BY-4.0, inherited from the source checkpoint. Model by Qiuqiang Kong et al. (PANNs). Core ML conversion by schism-audio.

Downloads last month: 19