Reducing Bias in Production Speech Models
Computation and Language
2017-05-15 v1
Abstract
Replacing hand-engineered pipelines with end-to-end deep learning systems has enabled strong results in applications like speech and object recognition. However, the causality and latency constraints of production systems put end-to-end speech models back into the underfitting regime and expose biases in the model that we show cannot be overcome by "scaling up", i.e., training bigger models on more data. In this work we systematically identify and address sources of bias, reducing error rates by up to 20% while remaining practical for deployment. We achieve this by utilizing improved neural architectures for streaming inference, solving optimization issues, and employing strategies that increase audio and label modelling versatility.
Cite
@article{arxiv.1705.04400,
title = {Reducing Bias in Production Speech Models},
author = {Eric Battenberg and Rewon Child and Adam Coates and Christopher Fougner and Yashesh Gaur and Jiaji Huang and Heewoo Jun and Ajay Kannan and Markus Kliegl and Atul Kumar and Hairong Liu and Vinay Rao and Sanjeev Satheesh and David Seetapun and Anuroop Sriram and Zhenyao Zhu},
journal= {arXiv preprint arXiv:1705.04400},
year = {2017}
}