TADA 3B audio model

TADA-3B-ml is a multilingual text-to-speech model from Hume AI released mid-March 2026. The full name is Text-Acoustic Dual-Alignment Large Language Model. It sits in the speech-language model group and runs on a Llama-3.2-3B base.

The model is open weight. Hume AI released both code and weights under an MIT license, so you can run it yourself if you want, although Llama's licence isn;t as permissive so it could be a bit of an issue. It’s free to use and built as a roughly 3B-class system, though the Hugging Face checkpoint shows about 4B parameters in BF16.

Hume describes TADA as a speech-language model rather than a classic TTS pipeline. Instead of generating long chains of tiny acoustic frames, it uses a token system where speech and text stay aligned. The goal is less lag and fewer spoken mistakes.

The checkpoint called HumeAI/tada-3b-ml is the multilingual version in the TADA family. Hume also released a smaller TADA-1B model plus a shared tada-codec encoder and decoder.

Here’s the basic idea. Most AI speech systems move text tokens and audio frames at different speeds. That mismatch can slow things down and sometimes cause wrong spoken words. TADA tries a different path – it lines text and speech up 1:1. One text token links to one speech representation. So the model steps through language and audio together… kind of a synchronized setup.

English works out of the box. The model also adds language aligners for several other languages. These include Arabic, Chinese, German, Spanish, French, Italian, Japanese, Polish and Portuguese.

One of the bigger claims around TADA is reliability. In its evaluation setup Hume says the system produced zero hallucinated words across more than 1,000 LibriTTSR samples. The team also reports a 0.09 real-time factor, which they say is over five times faster than similar LLM-based speech systems. Promising numbers… but still early-ish until more outside testing shows the same results.

Hume AI as a company mostly focuses on voice and emotion-aware systems. You might know their other tools like Octave for expressive speech, EVI – Empathic Voice Interface for live voice interaction, and their Expression Measurement models that analyze emotional signals. So TADA feels like part of a longer push into expressive voice tech rather than a one-off research drop.