Hume AI releases TADA-1B, an English TTS that aligns one speech vector per text token

Hume AI published TADA-1B, an English text-to-speech model built on Llama 3.2 1B. The Hub lists it at 2B parameters in BF16, though the name says 1B. The card describes 1:1 alignment, where the tokenizer produces one speech vector per text token, which it says removes the need for fixed-frame-rate processing, and dual-stream generation that produces a text token and the preceding token's speech together. Speech continuation is supported through a num_extra_steps setting.

The Evaluation section shows four charts (CER, real-time speed, speaker similarity, MOS naturalness) but the page text gives no numbers, so none are repeated here. The card mentions a separate TADA-3B-ml model sharing the same encoder. No Inference Provider hosts it, and the example code assumes a CUDA device.

The licence is tagged llama3.2, and the card says the project is built with Llama 3.2 under the Llama 3.2 Community License. The card gives no release date; the date used is the Hugging Face repo creation date, and the linked paper is listed as published 26 February.

An interesting architecture with its results in pictures.

Sources

• Hume AI model card on Hugging Face
• Hugging Face repo metadata (creation date)

Editor's first pass (AI-assisted)

• Real 3.5: Weights are public but you must run it yourself, from GitHub, on CUDA.
• Useful 3: English only, with no numeric results in the card text.
• Notable 3: A known voice-AI company opening a model.
• Novel 4: One speech vector per text token with dynamic duration is a distinct design.