Hume AI Launches TADA: An Ultra-Fast, Error-Free Speech Model

Hume AI Opens TADA to the Public

Hume AI recently announced the open-source release of TADA, an artificial intelligence model designed for speech generation. This system stands out for its ability to process text and audio in a synchronized manner. Unlike previous systems that generate significantly more audio frames per text token, TADA associates exactly one audio signal with each text token. This approach allows TADA to outperform its competitors in terms of speed, being more than five times faster.

In tests involving over 1,000 samples, TADA demonstrated impressive performance by producing no transcription hallucinations, meaning no invented or omitted words compared to the source text. In terms of naturalness, the system received a score of 3.78 out of 5 in human evaluations.

Compatibility and Accessibility

Hume AI claims that TADA is compact enough to be used on smartphones, although longer texts may sometimes result in a slight delay in voice output. The model is available in two sizes: 1 billion and 3 billion parameters, both based on Llama technology. The smaller version supports only English, while the larger version covers seven additional languages.

For those interested in technical details or wishing to contribute, all code and models are accessible on GitHub and Hugging Face under the MIT license.