Ten years of applied research into what human speech carries beyond its words, and into the systems required to keep that intact across a language boundary.
Machine translation has become extraordinarily good at the part of a message that can be written down. It remains almost entirely indifferent to the part that cannot.
Pitch, timing, stress, hesitation, register, the small refusals and concessions that a speaker never states outright. These are not decoration around the content. In most human exchanges they are a substantial share of the content. A listener acts on them before the sentence has finished. Strip them at the language boundary and two parties can receive faithful translations of the same words and leave with materially different understandings of what was agreed.
The stakes scale with the consequence of the conversation. In a negotiation, a diagnosis, a deposition, a de-escalation, the discarded layer is often the one that determined the outcome.
That gap is the company's research subject. Not translation quality; the field has that well in hand. The transmission of everything translation was never designed to carry.
Since 2025 the work has been concentrated in a single build, the RCG Voice Model. What it is for is described by the four questions below. How it is built is not published, and will not be. That boundary is set out on the research philosophy page.
What follows describes what the research is trying to establish and why it is difficult. It does not describe how. That distinction is deliberate and is explained below.
Establishing which components of a spoken message carry decision-relevant information, how much of that survives a conventional translation pipeline, and how the loss is distributed, because it is not distributed evenly. Some speakers, some languages and some registers lose far more than others, and a system that does not know this will fail hardest where it matters most.
Categorical affect classification, sorting an utterance into anger, joy, sadness, neutral, discards the gradient and the context that gave the signal its meaning. The research treats emotional content as continuous and situated, and as something carried with a message rather than re-inferred on the receiving side, where the evidence for it has already been destroyed.
Any system capable of carrying a speaker's affect is, by construction, capable of reproducing it. Consent, provenance, attribution and the limits of synthesis are treated here as design constraints inside the research rather than as policy applied afterwards. A boundary that is not built in is not a boundary.
The program's longest-range line: presence reconstructed from voice, from frequency and energy, rather than from captured image. The granted patent contemplates representations that carry adjusted delivery and culturally appropriate manner into the target language. This is where the work is going, and it is why the research outlived the product.
Across the platform years RealComm Global worked with independent engineering, language-technology and research partners: full-stack architecture and telecommunications engineering, multilingual speech and language technology, and research collaboration. Each was engaged for a defined scope of work under agreements that settled intellectual property boundaries in advance.
Those engagements have ended. Design and direction were always held in-house, which is why the intellectual property sits where it does, and the research continues without them. The company is open to new collaboration on the terms above.
Institutional researchers, language-technology groups and prospective licensees are all read by the founder directly.
