October 2025 Summaries

Krew, a company focused on building voice agents for credit servicing, faced significant challenges in ensuring compliance and reliability in one of the most regulated industries. To address these, they partnered with Coval, a platform specialized in voice AI evaluation, which offered rapid implementation and expert guidance, enabling Krew to conduct comprehensive testing in a matter of days instead of months. Coval's infrastructure allowed Krew to effectively manage perceived latency, turn-taking, and compliance metrics, significantly improving their evaluation process. This partnership not only enhanced Krew's testing capabilities but also provided third-party validation that bolstered their enterprise sales by demonstrating reliability and compliance to potential customers. Coval's role in transforming Krew’s QA strategy into a competitive advantage highlights the importance of rigorous, consistent testing in proving the reliability and compliance of voice AI systems in regulated sectors.

Deepgram's new transcription model, Flux, represents a significant advancement in Voice AI by addressing the long-standing challenge of balancing interruption and latency in conversational AI systems. Unlike traditional models that struggle with turn-taking due to reliance on silence detection or voice activity detection, Flux's streaming-first architecture and rapid updates enable seamless dialogue flow, akin to human conversation, without sacrificing accuracy. Benchmarks reveal that Flux significantly reduces latency to first token, maintaining equivalent word error rates to previous models like Nova-3, while handling end-of-turn transitions smoothly. This innovation is part of Deepgram’s larger Neuroplex architecture, which seeks to introduce a more context-aware approach by mimicking the brain's structure, thus enhancing voice AI with more lifelike agent behaviors and multi-dimensional context. This marks a new era for speech models, urging engineering leaders to consider the implications of real-time, streaming-native, and context-aware systems on their current architectures and processes.