G-Augment: Searching for the Meta-Structure of Data Augmentation Policies for ASR

For example, in [16], the authors discovered become automated and more "end-to-end," the data augmentation that SpecAugment [11] did not compose well with multi-style policy (what augmentation functions to use, and how training augmentation [4, 17], and found that they needed to to apply them) remains hand-crafted. We present G(raph)- ensemble the augmentations to benefit from both. Augment, a technique to define the augmentation space as In this work, we address this problem by a scheme we refer directed acyclic graphs (DAGs) and search over this space to as G(raph)-Augment, where a stochastic augmentation to optimize the augmentation policy itself. We show that policy is parameterized by a directed acyclic graph (DAG) given the same computational budget, policies produced by whose edges are labeled by sampling probabilities and augmentation G-Augment are able to perform better than SpecAugment parameters. By simultaneously searching for the policies obtained by random search on fine-tuning tasks on graph structure and the parameters that label the graph, we CHiME-6 and AMI. G-Augment is also able to establish are able to optimize not only the augmentation parameters of a new state-of-the-art ASR performance on the CHiME-6 the individual augmentations, but how those augmentations evaluation set (30.7% WER). We further demonstrate that are being applied. We utilize 17 ASR augmentations in our G-Augment policies show better transfer properties across search space, details of which can be found in section 3.3.