TADA by Hume AI

What is TADA by Hume AI?

TADA (Text-Acoustic Dual Alignment) is Hume AI's open speech model, released as a 1B English model and a 3B multilingual model, both built on Meta's Llama 3.2. Most LLM-based speech models generate dozens of audio frames per second; TADA pairs each text token with exactly one acoustic vector, so it needs far fewer steps and has no room to skip or add words.

It is Hume's only TADA release so far, and it is not sold through Hume's API: you download the weights and run them yourself, or try the free demo on Hugging Face. Hume says it is trained only to continue speech, so an assistant voice needs fine-tuning.

Who is TADA by Hume AI best for?

TADA by Hume AI is best for developers who host their own voice features, scripts that must be read word for word, long narration and speech research.

• Self-hosted voice features

  The weights run on your own GPU or Apple Silicon Mac, so there is no per-character bill and no audio has to leave your machine.

• Scripts read word for word

  The one-to-one text and audio alignment is built to stop skipped or invented words, and Hume measured none across more than 1,000 test clips.

• Long narration

  Hume estimates about 700 seconds of speech fit in a 2,048-token context, around ten times what a typical LLM-based voice model manages.

• Speech researchers

  The code, both model sizes, the audio tokenizer and the decoder are public, with a paper on arXiv, so you can fine-tune or extend it.

What's included with TADA by Hume AI?

One acoustic vector per text token, voice prompting from a short clip, 1B and 3B sizes, an Apple Silicon port, prompt caching and speech continuation.

One vector per text token

Each text token carries one acoustic vector, so a second of speech takes 2 to 3 steps instead of the 12.5 to 75 other systems use.

Voice prompting

Give it a short reference clip and its transcript, and it speaks your new text in that voice.

Two model sizes

TADA-1B covers English; TADA-3B-ML adds nine more languages, including Arabic, Japanese and Polish, through language-specific aligners.

Apple Silicon port

An MLX version runs on Apple Silicon Macs, and Hume says its 4-bit mode is about 10x faster and uses 60% less memory.

Prompt caching

Encode a reference voice once, save it and reuse it without loading the encoder, which saves about 2.5 GB of GPU memory.

Speech continuation

Without new text, it can keep talking after the prompt, writing its own words and speaking them in the same stream.

How does TADA by Hume AI work?

1. Accept the Llama licence

   Request access to Meta's Llama 3.2 1B or 3B on Hugging Face, since TADA uses its tokenizer.

2. Install the package

   Run pip install hume-tada, or pip install mlx-tada on an Apple Silicon Mac; the weights download from Hugging Face.

3. Encode a reference voice

   Load a short clip with its transcript; the encoder lines each word up with the audio, and you can save the result to reuse.

4. Generate speech

   Pass the prompt and your new text to the model, and it returns speech in the reference voice.

How much does TADA by Hume AI cost?

TADA by Hume AI is free: the weights and code cost nothing to download and the Hugging Face demo is free, so you pay only for the hardware you run it on.

Hume does not sell TADA through its API; its paid plans, from $3 a month, cover its Octave text-to-speech and EVI voice models instead. Running TADA costs only the hardware you run it on. You must accept Meta's Llama 3.2 licence on Hugging Face before the model will load. Full pricing on github.com. We checked these prices on 28 Sep 2026.

How do readers rate TADA by Hume AI?

No one has rated TADA by Hume AI yet. Be the first.

TADA by Hume AI pros and cons

It is fast for an LLM-based voice model and sticks to the script, but it is heavy on GPU memory, needs fine-tuning for assistant use and takes more setup than a hosted API.

What are the best TADA by Hume AI alternatives?

Three tools on Recaply do TADA by Hume AI's job. Chatterbox Turbo and Qwen3-TTS CustomVoice 1.7B are open source, so you can run them on your own server.

Chatterbox TurboOpen Source

Choose it when you want a smaller 350M model under a plain MIT licence, with tags such as [laugh] and [sigh]; every clip it makes carries Resemble's PerTh watermark.

Qwen3-TTS CustomVoice 1.7BOpen Source

Choose it when you want nine ready-made voices you steer with plain-language instructions in 10 languages, under an Apache 2.0 licence with no Llama sign-up.

Hume AIFree

Choose it when you would rather call a hosted API than run a model: its Octave voices include voice cloning, and commercial use starts on the $14-a-month Creator plan.

Questions about TADA by Hume AI