An ONNX export of Google's YAMNet, an audio event classifier trained on
AudioSet. It predicts 521 audio event
classes and also exposes a 1024-dimensional embedding that is well suited to
transfer learning on small, custom sound datasets.
This repository exists so that Audio Magic's
"Detect Sounds" tool can fetch the model from a host we control. It is a
straight format conversion — no weights were retrained, fine-tuned, pruned or
quantized.
Provenance
Upstream model
google/yamnet version 1 (formerly tfhub.dev/google/yamnet/1)
As required by Apache 2.0 §4(b), the modifications are:
The TensorFlow SavedModel was converted to ONNX with tf2onnx
(--saved-model, opset 15). No weight values were altered.
yamnet_class_map.csv is copied verbatim from the SavedModel's assets/
directory for convenience.
Nothing else was changed. The original NOTICE file is not reproduced because
upstream does not ship one.
Inputs and outputs
The mel-spectrogram frontend is baked into the graph, so the model consumes
a raw waveform directly — there is no feature extraction to reimplement (and so
no opportunity to get it subtly wrong).
Input — waveform, float32, shape [num_samples]: mono PCM at
16 kHz, nominally in [-1.0, 1.0]. Length is dynamic.
Outputs, one row per frame:
Name
Shape
Meaning
output_0
[frames, 521]
Per-class scores over the AudioSet ontology
output_1
[frames, 1024]
Embeddings (the useful part for transfer learning)
output_2
[frames × 96, 64]
Log-mel spectrogram patches
Each frame covers 0.96 s of audio and the window advances 0.48 s, so
frame i spans [i × 0.48, i × 0.48 + 0.96] seconds. Input shorter than one
patch is zero-padded and still yields one frame. Empirically:
1import numpy as np, onnxruntime as ort
23session = ort.InferenceSession("yamnet.onnx")4waveform = np.zeros(16000*3, dtype=np.float32)# 3 s of 16 kHz mono5scores, embeddings, log_mel = session.run(None,{"waveform": waveform})
Verification
The conversion was checked against YAMNet's documented reference behaviour
rather than assumed correct:
Input
Top class
1 kHz sine wave
Beep, bleep
Digital silence
Silence (1.00)
White noise
Static / White noise
Frame counts match the published 0.96 s / 0.48 s framing, and outputs were
compared numerically against an independent existing conversion across sine,
silence, noise, chirp and sub-second inputs.
Licence
Apache 2.0, inherited from the upstream model and code.
Copyright 2022 Google LLC. The full licence text is in LICENSE.
The AudioSet ontology and labels are released by Google under CC BY 4.0.
Citation
YAMNet implements the architecture described in:
bibtex
1@inproceedings{hershey2017cnn,
2 title = {{CNN} Architectures for Large-Scale Audio Classification},
3 author = {Hershey, Shawn and Chaudhuri, Sourish and Ellis, Daniel P. W. and
4 Gemmeke, Jort F. and Jansen, Aren and Moore, R. Channing and
5 Plakal, Manoj and Platt, Devin and Saurous, Rif A. and
6 Seybold, Bryan and Slaney, Malcolm and Weiss, Ron J. and
7 Wilson, Kevin},
8 booktitle = {2017 IEEE International Conference on Acoustics, Speech and
9 Signal Processing (ICASSP)},
10 year = {2017}
11}
1213@inproceedings{gemmeke2017audioset,
14 title = {Audio Set: An ontology and human-labeled dataset for audio events},
15 author = {Gemmeke, Jort F. and Ellis, Daniel P. W. and Freedman, Dylan and
16 Jansen, Aren and Lawrence, Wade and Moore, R. Channing and
17 Plakal, Manoj and Ritter, Marvin},
18 booktitle = {Proc. IEEE ICASSP 2017},
19 address = {New Orleans, LA},
20 year = {2017}
21}
It uses the MobileNetV1 depthwise-separable convolution architecture:
bibtex
1@article{howard2017mobilenets,
2 title = {{MobileNets}: Efficient Convolutional Neural Networks for Mobile
3 Vision Applications},
4 author = {Howard, Andrew G. and Zhu, Menglong and Chen, Bo and
5 Kalenichenko, Dmitry and Wang, Weijun and Weyand, Tobias and
6 Andreetto, Marco and Adam, Hartwig},
7 journal = {arXiv preprint arXiv:1704.04861},
8 year = {2017}
9}
