TADA

TADA is what happens when you actually solve the core problem instead of papering over it. Hume AI built Text-Acoustic Dual Alignment—a tokenization schema that synchronizes text and speech one-to-one inside the language model. Here's why that matters: a single second of audio contains 12.5 to 25 acoustic frames but only 2 to 3 text tokens. That mismatch is why every existing LLM-based TTS system is stuck choosing between speed, quality, and reliability. TADA eliminates the tradeoff by aligning audio directly to text—one continuous acoustic vector per text token—creating a synchronized stream where both move through the model in lockstep. The result: 5x faster speech generation than comparable systems, competitive voice quality, virtually zero content hallucinations, and a footprint light enough for on-device deployment. They're open-sourcing it because the future of voice AI shouldn't live behind a paywall.

Jakub Piotr Cłapa, Zach Krall (Head of Design), and Zachary Greathouse built this at Hume AI. The team tackled a fundamental mismatch that everyone else was working around by adding intermediate "semantic" tokens or reducing frame rates—both of which trade expressiveness for speed. TADA said no to those compromises and went straight to the source. This is now the fastest LLM-based TTS system available, and it's light enough to run on your phone. Voice AI just got a serious upgrade.