Streaming Sequence Transduction through Dynamic Compression

Explore a groundbreaking Transformer-based model called STAR (Stream Transduction with Anchor Representations) in this 46-minute lecture from the Center for Language & Speech Processing at Johns Hopkins University. Discover how STAR efficiently performs sequence-to-sequence transduction over streams by dynamically segmenting input streams and creating compressed anchor representations. Learn about its impressive performance in Automatic Speech Recognition (ASR), achieving nearly lossless compression at 12x and surpassing existing methods. Examine STAR's superior capabilities in simultaneous speech-to-text tasks, including improved segmentation and optimized latency-quality trade-offs. Gain insights into how this innovative model balances latency, memory footprint, and quality for streaming sequence transduction applications.