G-Augment: Searching for the Meta-Structure of Data Augmentation Policies for ASR
Not on the map. Why?
Abstract (arXiv)
Data augmentation is a ubiquitous technique used to provide robustness to automatic speech recognition (ASR) training. However, even as so much of the ASR training process has become automated and more "end-to-end", the data augmentation policy (what augmentation functions to use, and how to apply them) remains hand-crafted. We present Graph-Augment, a technique to define the augmentation space as directed acyclic graphs (DAGs) and search over this space to optimize the augmentation policy itself. We show that given the same computational budget, policies produced by G-Augment are able to perform better than SpecAugment policies obtained by random search on fine-tuning tasks on CHiME-6 and AMI. G-Augment is also able to establish a new state-of-the-art ASR performance on the CHiME-6 evaluation set (30.7% WER). We further demonstrate that G-Augment policies show better transfer properties across warm-start to cold-start training and model size compared to random-searched SpecAugment policies.
Key facts. Cited by 2 papers in the map (data through 13 Oct 2026). Builds on 26 of its 54 references that are in the map. First posted on arXiv 19 Oct 2022.
Builds on
26 of 54 references.
References · 26
- 2020wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representationscited by 3,870
- 2020Conformer: Convolution-augmented Transformer for Speech Recognitioncited by 1,844
- 2017Improved Regularization of Convolutional Neural Networks with Cutoutcited by 1,627
- 2019Common Voice: A Massively-Multilingual Speech Corpuscited by 1,366
- 2019SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognitioncited by 1,359
- 2019Self-training with Noisy Student improves ImageNet classificationcited by 1,306
- 2017Learning Transferable Architectures for Scalable Image Recognitioncited by 1,100
- 2018Regularized Evolution for Image Classifier Architecture Searchcited by 907
- 2018AutoAugment: Learning Augmentation Policies from Datacited by 730
- 2019Libri-Light: A Benchmark for ASR with Limited or No Supervisioncited by 464
- 2017Population Based Training of Neural Networkscited by 359
- 2019Fast AutoAugmentcited by 243
14 more
- 2018TED-LIUM 3: twice as much data and corpus repartition for experiments on speaker adaptationcited by 224
- 2019Population Based Augmentation: Efficient Learning of Augmentation Policy Schedulescited by 153
- 2020Improved Noisy Student Training for Automatic Speech Recognitioncited by 140
- 2019Learning Data Augmentation Strategies for Object Detectioncited by 139
- 2020AutoML-Zero: Evolving Machine Learning Algorithms From Scratchcited by 139
- 2020CHiME-6 Challenge:Tackling Multispeaker Speech Recognition for Unsegmented Recordingscited by 131
- 2017Learning to Compose Domain-Specific Transformations for Data Augmentationcited by 91
- 2021BigSSL: Exploring the Frontier of Large-Scale Semi-Supervised Learning for Automatic Speech Recognitioncited by 88
- 2021SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Networkcited by 83
- 2019SpecAugment on Large Scale Datasetscited by 60
- 2017Smart Augmentation - Learning an Optimal Data Augmentation Strategycited by 54
- 2020Improving 3D Object Detection through Progressive Population Based Augmentationcited by 41
- 2021Automated Deep Learning: Neural Architecture Search Is Not the Endcited by 13
- 2020SapAugment: Learning A Sample Adaptive Policy for Data Augmentationcited by 9
Built on by
Citing papers per year: 2025: 1; 2026: 1.
2 papers.