Evaluate before you ship: introducing the Voice Live Evaluation Harness (opens in new tab)

You've built a voice agent on Azure Voice Live. It demos beautifully. Then a teammate asks the question that keeps every voice-agent team up at night: "How do we know it's actually good — across 200 customer calls, not the three we just listened to?" Until today, the honest answer was: put on headphones. Manual listening. Subjective scoring in a spreadsheet. No baseline, no regression signal, no way to defend a model swap with data. We're releasing the Voice Live Evaluation Harness to change ...