How Comcast Powers AI Speech Recognition with Weak Supervision

Explore the innovative approach to automatic speech recognition (ASR) developed by Comcast Applied AI in this 28-minute presentation. Learn how Lead Research Scientist Raphael Tang and his team overcame the challenges of limited labeled data and computational resources to create a highly efficient ASR system. Discover the use of weak supervision techniques, including leveraging a third-party ASR system and Snorkel labeling functions derived from implicit user feedback. Understand the novel inference acceleration method involving CUDA graphs of varying input lengths. Gain insights into the impressive results achieved, including an 8% relative improvement in word-error rate and a 600% speedup compared to third-party ASR systems. See how this groundbreaking system, named SpeechNet, now handles 12 million daily queries for voice-enabled smart televisions at Comcast.