Leveraging Pre-training Models for Speech Processing

Explore the cutting-edge advancements in speech processing through pre-training models in this comprehensive 2-hour and 59-minute lecture from the Center for Language & Speech Processing (CLSP) at Johns Hopkins University. Delve into the crucial role of pre-training in advancing speech, natural language processing (NLP), and computer vision (CV) research. Discover how pre-trained models, leveraging vast amounts of unlabeled data, can be transferred to multiple downstream applications with remarkable efficiency. Examine benchmarks like SUPERB, LeBenchmark, and NOSS, which evaluate pre-trained models' generalizability across various speech and audio processing tasks. Learn how small players in the field can benefit from publicly available pre-trained models, achieving impressive performance with minimal additional training. Investigate research directions including modeling prosody, efficient leveraging of pre-trained models in downstream tasks, robustness under domain mismatch, incorporation of visual information, and development of computationally efficient models. Gain insights into open-source benchmarks, tools, and techniques that drive the frontier of network pre-training technology in speech processing.

Leveraging Pre-training Models for Speech Processing