Hume AI open-sources TADA, a voice model that runs on your phone

TADA (Text-Acoustic Dual Alignment) is an open-source speech-language model from Hume AI designed to solve the clunky trade-offs of voice AI. Instead of juggling thousands of acoustic frames, it locks text tokens and continuous acoustic vectors into a strict one-to-one sync, delivering high-speed TTS without skipping words or making up sounds.

The codebase and pre-trained weights are available now. It supports 10 languages including Chinese, and it is compact enough to run locally on mobile phones and edge devices without touching a cloud API.

How it eliminates hallucinations

Standard LLM-based TTS models face an awkward balance: a single second of speech translates to just 2 or 3 text tokens, but demands anywhere from 12.5 to 75 acoustic frames. That flood of audio tokens bloats the context window, jacks up memory usage, and frequently leads the model to drift, skip phrases, or hallucinate words.

TADA bypasses discrete audio frames altogether. It maps audio representations directly to individual text tokens, assigning one continuous acoustic vector to each text token. Both modalities march in lockstep through the language model.

On input, an encoder and aligner pull acoustic features for each token. On output, the hidden state conditions a flow-matching head that decodes back into raw audio. Because the architecture enforces a rigid one-to-one mapping, the model cannot skip words or invent hallucinated sentences by design.

Speed, footprint, and current limits

In benchmarks, TADA clocks a real-time factor (RTF) of 0.09, outpacing similar-grade LLM TTS systems by more than 5x. Because it operates at just 2 to 3 frames per second of audio, inference is remarkably light.

Across more than 1,000 test samples on the LibriTTSR benchmark, TADA produced zero hallucinations under a character error rate threshold of 0.15. It also provides simultaneous transcription as it speaks, dropping the need for a secondary speech-to-text pipeline.

Context efficiency is another massive leap. While a traditional setup burns through a 2,048-token context window in about 70 seconds of speech, TADA holds roughly 700 seconds in that exact same budget. In human evaluations on the EARS dataset, it scored 4.18/5.0 on speaker similarity and 3.78/5.0 on naturalness, ranking second overall.

It is not without flaws, though. Over very long generations, the voice can experience speaker drift. Hume AI uses online rejection sampling to suppress this, but recommends periodically resetting the context window as a practical workaround.

Why it matters

If you build on-device voice assistants, mobile tools, or interactive hardware, TADA deserves an immediate look. It drags frontier-level conversational voice off expensive cloud clusters and drops it directly onto local chipsets, eliminating latency, cloud API bills, and privacy exposure in one stroke.

We think TADA's architectural bet is spot on. Rather than masking hallucinations with endless post-training or piling on complex discrete tokenizers, it tackles sequence mismatch at the root. For sensitive deployments in healthcare or customer support where a stray word is disastrous, that built-in reliability is a major advantage.