Transcriptor sound models (ONNX)
ONNX exports of three AudioSet sound taggers, used by the Transcriptor desktop app to find sounds (shots, laughs, music…) in a recording. They run locally with ONNX Runtime; nothing is sent anywhere.
Each graph takes waveform (float32, shape (batch, samples), mono, values in [-1, 1]) and returns probs (float32, (batch, classes), sigmoid scores). The log-mel front end is inside the graph. Opset 17, dynamic batch.
| File | Model | Rate | Window | Classes | SHA-256 |
|---|---|---|---|---|---|
yamnet.onnx |
YAMNet (Google) | 16 kHz | 15 600 samples (0.975 s) | 521 (YAMNet class map) | 98936f40391a8817375cc430d83060b2f63063410d1d8b6ca39658234a2a2bef |
efficientat-mn04-3s.onnx |
EfficientAT mn04_as (Florian Schmid) | 32 kHz | 96 000 samples (3 s) | 527 (AudioSet) | db3ae00d161763be541d2af53aa72c44f42cd9bcd7bd63d652aaa684fb436749 |
panns-cnn14-3s.onnx |
PANNs CNN14 (Kong et al.) | 32 kHz | 96 000 samples (3 s) | 527 (AudioSet) | f8d75f78a2d3884e8e1f8976d9e63d8c80752dc242b377e6fa0282988d94cb0e |
How they were made
- YAMNet: ported from the official Keras weights
yamnet.h5(tensorflow/models,research/audioset/yamnet) to PyTorch, then exported. Checked against TensorFlow 2.21's own YAMNet on 377 windows of speech and game sound: largest score difference 1.6e-5, same top class on every window. - EfficientAT mn04_as (release v0.0.1,
mn04_as_mAP_432.pt) and PANNs CNN14 (Cnn14_mAP=0.431.pth, Zenodo record 3987831): exported from the original PyTorch code; largest difference to it 2e-6.
Licences and attribution
yamnet.onnx: Apache-2.0. Derived from YAMNet, Copyright The TensorFlow Authors (tensorflow/models). Changes: converted from Keras weights to ONNX with the front end in the graph. Licence:licences/yamnet-tensorflow-models-LICENSE.txt.efficientat-mn04-3s.onnx: MIT, Copyright (c) 2022 Florian Schmid (github.com/fschmid56/EfficientAT). Licence:licences/efficientat-LICENSE.txt.panns-cnn14-3s.onnx: weights CC BY 4.0 (zenodo.org/records/3987831), code MIT (github.com/qiuqiangkong/audioset_tagging_cnn,licences/panns-audioset_tagging_cnn-LICENSE.MIT.txt). Attribution: Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, Mark D. Plumbley, "PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition", IEEE/ACM Transactions on Audio, Speech, and Language Processing 28 (2020). Changes: converted to ONNX with the front end in the graph.
All three are trained on AudioSet (YouTube clips).