Link to paper: https://aclanthology.org/2021.konvens-1.19.pdf
Link to talk: https://youtu.be/xYqrwV2y2jc, Slides in this repository.
ACL Anthology: https://aclanthology.org/2021.konvens-1.19/
BibTeX:
@inproceedings{eberhard-zesch-2021-effects,
title = {Effects of Layer Freezing on Transferring a Speech Recognition System to Under-resourced Languages},
author = {Eberhard, Onno and Zesch, Torsten},
booktitle = {Proceedings of the 17th Conference on Natural Language Processing (KONVENS 2021)},
month = "6--9 " # sep,
year = {2021},
address = {D{\"u}sseldorf, Germany},
publisher = {KONVENS 2021 Organizers},
url = {https://aclanthology.org/2021.konvens-1.19},
pages = {208--212}
}
This repository contains additional resources for the paper "Effects of Layer Freezing on Transferring a Speech Recognition System to Under-resourced Languages".
The steps taken for training the models are compiled in the file training.sh. The training and testing output logs can be found in the logs directory.
The modified versions of DeepSpeech that utilize layer freezing can be found at https://github.com/onnoeberhard/deepspeech-transfer, the different versions with the first N layers frozen are in the branches transfer-1 (N = 1), transfer-2 (N = 2), transfer (N = 3), transfer-lstm (N = 4) and transfer-5 (N = 5). The version where the fifth instead of the LSTM layer is frozen is in branch transfer-4.
For freezing the LSTM, some custom modification of TensorFlow had to be made. This version of DeepSpeech uses TensorFlow 1.15.2 and the LSTM (when training on GPU) is an instance of CudnnLSTM. To make this class freezable, we had to modify the method _create_saveable of the _CudnnRNN class by exchanging all instances of self.trainable_variables by self.trainable_variables if self.trainable else self.non_trainable_variables.