Deepgram's VP of Research Andrew Seagraves joins to explore the science and engineering behind modern speech recognition systems. Hermes and Andrew dive deep into why speech recognition isn't a solved problem, the two-stage training process of speech-to-text models, and the challenges of balancing real-time latency with accuracy. The conversation covers Deepgram's origins from dark matter research, power laws in speech data, buffer-based architectures for real-time transcription, and frontier challenges like multilingual code-switching, emotion detection, and conversational dynamics. Andrew shares insights on model deployment, customer use cases from NASA to food ordering, and the future of self-adapting speech models.

Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world