schism-audio/cnn14-audioset-mlx

PANNs CNN14 (AudioSet) — MLX

PANNs CNN14 (Cnn14_mAP=0.431, Kong et al. 2020, 527-class AudioSet tagging) converted to MLX for Apple Silicon. Source checkpoint: Zenodo record 3987831 (original implementation).

Runs with the open schism-mlx package:

from schism_mlx.classify.cnn14 import Cnn14

model = Cnn14.from_pretrained("schism-audio/cnn14-audioset-mlx")
for label, score in model.classify("song.wav", top_k=5):
    print(f"{score:.3f}  {label}")

Unlike AST, CNN14 handles arbitrary-length audio — no 10-second truncation. ~7 ms per 10-second clip on an M5 Max (RTF 0.0007).

Conversion & verification

• Weights are the original float32 checkpoint, tensor-renamed and with convs permuted to MLX's channels-last layout. No quantization, no finetuning.
• Verified end to end against a torch reference reconstructed from the released checkpoint's own frozen STFT/mel buffers: max logit difference 2.9e-6, identical top-5 on synthetic signals and real drum recordings from schism-audio/e-gmd.
• The librosa-style log-mel frontend (32 kHz, 64 mel bins, 50–14000 Hz) is reimplemented in pure numpy; its mel filterbank matches the checkpoint's frozen melW bit-exactly. No torch/librosa needed at inference time.

Precision note: on M5-class GPUs, MLX silently runs float32 matmuls at TF32-class precision (ml-explore/mlx#3860). schism_mlx disables this on import; if you load these weights with your own MLX code, set MLX_ENABLE_TF32=0 to match the reference implementation.

License

CC-BY-4.0, inherited from the source checkpoint. Model by Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D. Plumbley (PANNs); MLX conversion by schism-audio.