Language is the most successful technology humans ever invented for compressing the world. We took experience i.e., every smell, every sound, every face, every hesitation before an answer.. and found a way to shrink it into discrete symbols that fit in a paragraph. It works so well that we forget it happened and it was well.. lossy.
We mistake the map for the territory.
Voice AI inherited that mistake. The models that power the Voice AI systems, even today, were built on one implicit assumption: Massive amounts of clean, transcribed, single-language audio data would always be available. And when you have large models at the backbone, you have to follow such laws of scale. For English, this is fair. For three quarters of people alive today, that is unfair - not linguistically, not phonetically!
The consequences have led us substandard speech models. English speakers get production-grade, natural-sounding, emotionally expressive voice AI. Everyone else gets semi-robotic, accent-mangled, code-switch-breaking approximations we cannot deploy without embarrassing themselves in front of our own customers. When the world majorly moves towards more and more B2C voice applications (when consumer AI apps explode in late 2027), this problem will amplify.
For last 4 years, AI's most spectacular results have come from learning to predict the next compressed symbol. The whole internet, tokenized. Out of that simple objective came models that can write essays, debug code, and pass the bar. Speech, Visuals, Videos and most of the rendering for the real-world is where that runs out.
Your voice is warmer when you talk to someone you love. It tightens when you lie. It slows when you're sad. The room you're in is in the recording. So is the time of day, the device, the language you switched to when the English word didn't quite fit. None of this is incidental. All of it has meaning. This is what makes us human.
So we stopped predicting words and started predicting the signal, in a latent space wide enough to hold what a transcript throws away. That got us halfway. A model trained on thin data is uncertain constantly, and what it does with that uncertainty decides whether it sounds human or not. And the output which you see breaking with every other speech model out there - is worst exactly where the data is thinnest, which is every language we actually care about.
We think the foundation needed a breakthrough. Our fix was to change what the network reports. Not a guess, but the evidence behind the guessed utterance outputs. We wrote the whole thing up, derivations and ablations and an honest ledger of what we measured against what we still owe, in our experiements. You can read about it here →
Speech intelligence needs to be built the way intelligence seems to actually work: not by predicting what was said, but by predicting what the world is like - in a latent space rich enough to hold the things words throw away at a fraction of the training cost. What you get from all of this is the new architecture, not a fine-tuned layer that sits on top of it. The evidential encoder basically learns the structure of the speech and its own confidence in that structure at the same time and the latent space it builds is where the meaning sits by continual tokenizer than standard discrete ones which throw the speech metadata away. Our state-of-the-art DCASE 2026 winning domain agnostic leaning algorithms align that latent space across domains and languages - so you bypass the catastrophic forgetting i.e., learning every new language or dialect, you do not need to keep forgetting the older weights. Together they make up layerwise joint embedding predictive architecture.
That's what we're building: A world model for speech. A foundation that learns meaning in a continuous latent space, trained on the most of world's dialects, accents and slangs. Most of the world’s languages have minimal to no speech data to have SOTA models on, thus our foundation model - both compute and data efficient can form a speech AI backbone for the mankind.
