Hume AI has released TADA (Text-Acoustic Dual-Alignment), an open-source text-to-speech model that eliminates a persistent problem in AI speech synthesis: hallucinated words. Built on Meta's Llama 3.2 1B architecture, TADA uses a 1:1 token alignment system where every audio token maps directly to a text token, making it structurally impossible for the model to generate words that are not in the input transcript.

What Happened

On March 10, 2026, Hume AI published TADA on GitHub and HuggingFace under the MIT license. The model is a unified speech-language system with 2 billion total parameters that generates speech in a single forward pass. Two versions are available: tada-1b for English and tada-3b-ml, a 4-billion-parameter multilingual variant. The accompanying research paper is available on arXiv.

TADA's core innovation is its dual-alignment architecture. Unlike conventional TTS models that generate audio tokens in a sequence loosely tied to text, TADA enforces strict 1:1 correspondence between text and audio tokens. Each text token produces exactly one audio token, and the model handles prosody and duration synthesis per token in a single step through what the team calls dynamic duration synthesis.

Why It Matters

Transcript hallucination has been one of the most frustrating issues in neural TTS. Models sometimes insert, skip, or substitute words during speech generation, producing audio that does not match the input text. This is a dealbreaker for accessibility applications, audiobook production, and any use case where accuracy is non-negotiable.

TADA solves this architecturally rather than through post-processing filters or alignment corrections. Because the 1:1 token mapping is baked into the model structure, there is zero chance of the output containing words that were not in the input. The team reports that TADA also runs approximately 5x faster than comparable open-source TTS models, which matters for real-time applications and cost-sensitive deployments.

The MIT license makes TADA genuinely open source with no commercial restrictions. This stands in contrast to several recent TTS releases that use more restrictive licenses. For developers building on Llama-based architectures, the familiar foundation lowers the barrier to fine-tuning and integration. Teams already working with open-source speech tools like Fish Audio's S2 model now have another strong option in the pipeline.

Key Details

  • Built on Meta's Llama 3.2 1B with 2 billion total parameters
  • 1:1 token alignment ensures zero transcript hallucination
  • Dynamic duration synthesis handles prosody and timing in a single step
  • Two variants: tada-1b (English) and tada-3b-ml (multilingual, 4B parameters)
  • MIT license with no commercial restrictions
  • Approximately 5x faster than comparable open-source TTS systems
  • Research paper: arXiv:2602.23068

What to Do Next

Developers working on speech synthesis should clone the TADA repository and test it against their current TTS pipeline. The tada-1b model is small enough to run on consumer hardware, making local evaluation straightforward. For multilingual projects, the tada-3b-ml variant supports multiple languages with the same zero-hallucination guarantee.

Content creators producing voiceovers, podcasts, or narration should pay attention to the accuracy claims. If TADA delivers on its zero-hallucination promise in practice, it removes the need for manual transcript verification that currently adds time to every TTS workflow. Test it with your specific content to see how it handles domain-specific terminology and pacing requirements.