Systems and methods for principled bias reduction in production speech models

US10657955B2 · kind B2 · utility

Assignee

• BAIDU USA LLC · US

Inventors

• Eric Dean Battenberg · Sunnyvale, US
• Rewon Child · Emeryville, US
• Adam Coates · Mountain View, US
• Christopher Fougner · London, GB
• Yashesh GAUR · Redmond, US
• Jiaji Huang · San Jose, US
• Heewoo Jun · San Francisco, US
• Ajay Kannan · Dublin, US
• Markus Kliegl · Santa Clara, US
• Atul Kumar · Palo Alto, US
• Hairong Liu · San Jose, US
• Vinay Vishwanath Rao · Mountain View, US
• Sanjeev Satheesh · Sunnyvale, US
• David Seetapun · Berkeley, US
• Anuroop Sriram · Sunnyvale, US
• Zhenyao Zhu · Mountain View, US

Key dates

• Filing date: Jan 30, 2018
• Grant date: May 19, 2020
• Expiry date: Jul 21, 2038

Classification

• Technology area (CPC G): Physics
• CPC primary: G10L25/18
• WIPO field: Computer technology
• WIPO sector: Electrical engineering

Abstract

Described herein are systems and methods to identify and address sources of bias in an end-to-end speech model. In one or more embodiments, the end-to-end model may be a recurrent neural network with two 2D-convolutional input layers, followed by multiple bidirectional recurrent layers and one fully connected layer before a softmax layer. In one or more embodiments, the network is trained end-to-end using the CTC loss function to directly predict sequences of characters from log spectrograms of audio. With optimized recurrent layers and training together with alignment information, some unwanted bias induced by using purely forward only recurrences may be removed in a deployed model.