AI voices sound entirely human. What's next? | Inworld TTS-2

In 2021, human listeners in the Blizzard Challenge could no longer tell AI-generated speech from a real voice actor. That's when the classic text-to-speech benchmark saturated. Since then, the field has moved to a harder target: expressiveness. This video is a deep dive into how voice AI labs define, measure, and train for expressive speech. It covers the mean opinion score (MOS), LLM-style arenas and Elo rankings for TTS, AI judges like UTMOS and DNSMOS, whether speech language models (Gemini, GPT Audio, Kimi Audio) can replace human raters, deterministic signal-processing proxies for expres