androiddrew's picture
ONNX release v1
e38c9f3 verified
Raw History Blame Contribute Delete
1.38 kB
Google speech_embedding/1 (TensorFlow1 / speech-embedding / version 1)
Original work published by Google under the Apache License, Version 2.0.
Source: https://www.kaggle.com/models/google/speech-embedding/TensorFlow1/speech-embedding/1
Original archive SHA-256:
0f626c42009c1eb916401b7b2800229a211e43100115bc591b9afbf1b60da27f
Paper: James Lin, Kevin Kilgour, Dominik Roblek, Matthew Sharifi,
"Training Keyword Spotters with Limited and Synthesized Speech Data",
arXiv:2002.01322. https://arxiv.org/abs/2002.01322
This repository contains an ONNX conversion made by Drew Bednar for
go-phonmatchnet, from the PyTorch conversion
androiddrew/google-speech-embedding-pytorch (revision v1, commit
ad7a3dd93b02f83797a040857425d6343d98680a). It is not affiliated with or
endorsed by Google.
Modifications to the original work:
- the TensorFlow graph was re-implemented with PyTorch operations, then
exported to ONNX;
- HWIO convolution kernels were transposed to OIHW;
- the original inference-time spatial-mean subtraction was applied to the
first convolution kernel ahead of time;
- the FFT was expressed as a convolution with a fixed DFT basis, with the
original window applied to it.
All trained parameters, normalization statistics and frontend constants are
otherwise unchanged. No training or fine-tuning was performed.
The original archive contained no NOTICE file.