MLCommons’ David Kanter, NVIDIA’s Daniel Galvez on Publicly Accessible Datasets - Ep. 167

Cost-Effective Labeling

Cost-effective labeling has revolutionized the way datasets are prepared for machine learning, significantly reducing costs and increasing accessibility. David Kanter explains that by using semi-supervised learning and computational power, the cost of labeling datasets was reduced from an estimated $5 million to just $3,000 1. This approach allows researchers to focus resources on high-quality expert labeling for critical cases or to kickstart new projects with minimal investment. Daniel Galvez highlights the use of a CUDA-based decoder for speech recognition, which accelerates the process by a factor of 1000, enabling the alignment of 50,000 hours of audio in just two days 2.

   

Speech Recognition Advancements

Advancements in speech recognition technology have significantly improved processing speeds and broadened accessibility. Daniel Galvez notes that the CUDA-based decoder developed by NVIDIA allows for rapid processing of large audio datasets, transforming the landscape of speech recognition tasks 2. This technology not only enhances efficiency but also supports diverse linguistic communities by including languages like Catalan in the multilingual spoken words corpus. David Kanter emphasizes the importance of such datasets, which include 350,000 keywords in 50 languages, impacting over 5 billion people globally 3.