Reproduce “State-of-the-art Speech Recognition With Sequence-to-Sequence Models” on OpenResearch

About this paper

Chung-Cheng Chiu, Tara N. Sainath, Yonghui Wu, Rohit Prabhavalkar, Patrick Nguyen, Zhifeng Chen, Anjuli Kannan, Ron J. Weiss and 6 more

Researchers at Google developed an enhanced Listen, Attend, and Spell (LAS) sequence-to-sequence model for automatic speech recognition. Their approach, integrating Word Piece Models, Multi-Head Attention, and advanced optimization techniques like Minimum Word Error Rate training, achieved a 16% relative Word Error Rate reduction on a large-scale voice search task compared to Google's conventional production system.

Key insights

• End-to-end Listen, Attend, and Spell (LAS) models, when enhanced with specific architectural and training improvements, can outperform complex conventional ASR systems on large-scale, real-world tasks.
• The combination of structural advancements like Word Piece Models (WPM) and Multi-Head Attention (MHA) with optimization techniques such as Minimum Word Error Rate (MWER) training, Scheduled Sampling, and Label Smoothing are crucial for achieving state-of-the-art E2E ASR performance.
• E2E models offer significant simplification and a much smaller first-pass model size for the core recognition engine compared to multi-component systems, demonstrating potential for more efficient deployment.

Summary by alphaXiv.