TADA 3B ML

Overview

TADA 3B ML is the multilingual checkpoint of TADA (Text-Acoustic Dual Alignment), Hume AI's open speech-language model. It sits on Llama 3.2 3B and uses the same audio encoder as the English-only 1B release.

Most other LLM-based speech systems run audio at a fixed frame rate, which Hume puts at 12.5 to 75 audio tokens for a second of speech that holds two or three text tokens. TADA gives each text token exactly one acoustic vector. Each autoregressive step handles one token and sets its duration and prosody, and a flow-matching head conditioned on the model's hidden state produces acoustic features, which are then decoded into audio. Hume says this one-to-one mapping means the model cannot skip or invent words.

What you can build

• Narration for e-learning modules or articles in a voice you supply, where every word of the script has to be spoken.
• Voice-overs in Spanish, German, Japanese, Arabic and the other listed languages, each cloned from a clip recorded in that language.
• Apps with saved voices: encode a speaker once, then send only the embedding file afterwards.

Inputs

• Prompt (required): the text to speak.
• Reference audio: one clip of the voice to clone (MP3, WAV, M4A, WEBM or OPUS), ideally 10 to 30 seconds. Treat it as required: the model needs a voice prompt unless you pass a saved embedding.
• Reference text: the clip's transcript. The encoder's built-in speech recognition is English-only, so other languages need it typed in.
• Language: English, Arabic, Chinese, German, Spanish, French, Italian, Japanese, Polish or Portuguese. Choose the one spoken in the reference clip.
• Embedding speaker: a .pt file from an earlier run, used instead of the clip.
• Save embedding speaker: True returns that .pt file with the audio. Default False.
• Max extra steps: 0 to 200, default 0. Above 0, the end-of-text marker after your script is removed and the voice keeps talking past it, which Hume calls speech continuation. In one test, 60 extra steps added about 15 seconds of audio. Keep it at 0 when the clip should stop at the last word.
• Seed: an integer, default 0, for repeatable output.

Outputs

An audio file of the spoken text (the sample on this page is an MP3). With Save embedding speaker on, you also get the encoded voice prompt as a .pt file.

Recommended settings

• Use a reference clip in the target language and paste its exact transcript, English included. For other languages, Hume notes that generation still works without one, but alignment quality degrades.
• Set Language to match the clip.
• Leave Flow matching steps at 10. Hume lowered the default from 20 and reports no perceptible quality loss.

Limitations

• Hume reports occasional speaker drift in long generations and suggests resetting the context, so split long scripts across several runs.
• Hume pre-trained the model on speech continuation only and says assistant-style use needs further fine-tuning.
• Aligners exist only for the ten languages in the Language menu.

Safety & compliance

The weights are released under the Llama 3.2 Community License, and use is subject to its terms. Clone a voice only with the consent of the person it belongs to.