Boson AI’s Higgs Realtime scores higher than GPT and Gemini in interruption recovery

Boson’s approach contrasts with traditional pipelines (speech recognition → language model → text-to-speech) by treating audio as a unified input/output stream. Co-founder Alex Smola, who previously led AI initiatives at AWS and researched at Carnegie Mellon University, emphasizes inference economics and multilingual consistency as key differentiators. The company’s ProactBench research explores proactive assistance (e.g., suggesting remote appointments if a user mentions mobility limitations), but Smola stresses balancing anticipation with user control. Boson claims its system handles thousands of concurrent calls for a global insurance company at one-third the cost of human agents, with conversation quality improving from 30 to 80 over five months—far exceeding GPT-4o Realtime’s 55 on the same metric.