Year One Was All Psychology, Not Code: Inside Futuro's First Year of Human Conversation Research

Most AI startups launch with a model and a marketing budget. Futuro launched with a research team and zero engineers. In September 2021 — well before "conversational AI" was a category anyone outside the research community recognized — we set out to answer one question: what does the human brain actually use to decide whether the voice on the other end of a call is human or not? That question turned out to be much harder than we'd anticipated, and the answer turned out to be much weirder. The first year of the company was spent entirely on the research that would, four years later, produce the 94% human indistinguishability rate we measured in our 2025 double-blind study. This is the story of that year — what the research looked like, who ran it, what we learned, and why the findings made every existing AI voice technology unsalvageable.

The Counterintuitive Choice — No Engineers in Year One

The original Futuro org chart from late 2021 had nine people on it. Seven were researchers. Two were operations. Zero were engineers. That ratio was deliberate, and it was the single most important early decision the company made.

The reasoning was uncomfortable: every AI voice startup we'd looked at had launched in roughly the same way. Pick an off-the-shelf text-to-speech provider, wrap it in a prompt, build a UI, raise money, ship to market. The differentiator was usually "we have a slightly better prompt" or "we integrate with one more CRM than they do." None of them sounded like real humans on a phone call, and none of them seemed to be asking why.

We made an early bet that the why mattered more than anything else. Before writing a single line of code, before evaluating any TTS vendor, before committing to any specific model architecture, we wanted to understand — in detail — what the human brain actually uses to distinguish a human voice from a synthetic one. That was a research project, not an engineering project. It needed cognitive scientists and linguists and speech pathologists, not Python developers.

So we hired the research team first.

The Research Question

The question we set out to answer was deceptively simple: what does the human brain actually use to decide whether the voice on the other end of a call is human or not?

The simplicity was deceptive because every obvious answer was wrong. "It's about pronunciation" — no, modern TTS pronounces words almost flawlessly and listeners still detect AI within seconds. "It's about tone" — also no, AI voices have nuanced tone control. "It's about word choice" — definitely not, the LLMs of 2021 could already write convincingly human-sounding text. "It's about the topic" — no again, listeners detected AI in our blind tests regardless of subject matter.

Whatever the brain was using, it was something subtler than any of these surface features. And the only way to find out what it was — given that the conversational-AI research community in 2021 was tiny and the existing literature was thin — was to run our own studies, with our own protocols, designed specifically to surface the answer.

Why 2021 Had No Useful Data

To put the timing in context: ChatGPT had not yet been released. Voice cloning was a research curiosity, not a deployable technology. The phrase "conversational AI" was used mostly by chatbot vendors, and those chatbots were almost universally text-only. The few voice products that existed (Alexa, Siri, Google Assistant) were optimized for short command-and-response interactions, not multi-turn conversation.

The academic literature on synthetic-voice perception was small and largely focused on TTS pronunciation accuracy — measuring whether listeners could understand what was being said, not whether they could detect that it was being said by a machine. There was rich literature on phonetics, prosody, and speech disfluencies in human conversation, but almost no work bridging that to AI voice perception specifically.

We were going to have to do the bridging work ourselves. That meant:

• Designing listening-test protocols that could surface unconscious detection cues
• Recruiting participants in volume large enough to get statistically meaningful results
• Pairing human and synthetic samples carefully enough to isolate the variables we wanted to measure
• Building the analytical apparatus to interpret what listeners were responding to even when they couldn't articulate it

This was a 12-month research effort, not a six-week sprint.

The Team We Assembled

The seven research hires represented four disciplines that don't normally work together:

Cognitive psychology. Two PhDs whose work had focused on auditory attention and unconscious pattern detection. Their job was to design protocols that surfaced what listeners were responding to even when listeners themselves couldn't explain it.

Linguistics and phonetics. Two researchers with deep backgrounds in conversational analysis and prosodic patterning. They'd spent careers cataloging the micro-features of human speech — the breath patterns, the disfluency timing, the way intonation shifts mid-sentence — that everyone uses without noticing.

Speech pathology. One clinician whose practice had focused on communicative disorders. The relevance: she'd spent twenty years studying what happens when speech goes wrong, which turned out to be the inverse of the question we needed to answer (what does it look like when speech is convincingly normal?).

Acoustic engineering. Two engineers who could build the technical apparatus the research team needed — sample libraries, blind-test infrastructure, audio analysis tools — without committing us to any specific product architecture.

The four disciplines fed into each other in ways we didn't anticipate. The cognitive psychologists would surface a finding ("listeners detect AI within 800ms but can't articulate why"), the linguists would translate it into testable hypotheses about specific speech features, the speech pathologist would point out which of those features mapped to known clinical patterns, and the acoustic engineers would build the tools to test the hypotheses at scale.

The Methodology — Listening Tests at Scale

The core experimental protocol stayed roughly the same across the year, with refinements as we learned more.

A listener — recruited from a pool of paid participants screened to balance age, regional background, and prior AI exposure — would be played a sequence of audio samples. Each sample was a 30-90 second segment of conversation. Some were recordings of real human-to-human phone calls. Some were synthetic samples produced by then-state-of-the-art TTS systems. Some were hybrids — human recordings with specific features systematically removed (breath sounds suppressed, filler words edited out, prosody flattened) — so we could isolate which features mattered.

After each sample, the listener was asked one question: was this voice a real human or a machine? Yes/no. Optional follow-up: what tipped you off? Most listeners couldn't articulate an answer to the follow-up.

By the end of year one, we had run roughly 14,000 individual listener-sample evaluations across approximately 2,200 participants. The dataset was large enough to surface signals that any individual listener couldn't articulate but the aggregate clearly showed.

The Six Categories of Human-Detection Signals

The findings clustered into six distinct categories of cues that the human brain used to detect AI voices. None of them, individually, were obvious. All of them, collectively, explained essentially the entire variance in our detection-rate data.

1. Disfluency density. Real human speech has roughly 6-12 disfluencies per minute — "umm," "you know," "right," "so," self-corrections, repeated words. Listeners detected AI when this count dropped below 2-3 per minute, even when nothing else about the speech was unusual.

2. Breath patterning. Humans inhale before long sentences, exhale slightly during transitions, and audibly breathe at predictable points. AI samples that had been engineered to be "smooth" had no breath sounds at all. Listeners flagged the absence with high reliability — even though they almost never said "no breaths" when asked what tipped them off.

3. Prosodic micro-variance. Even within a single declarative sentence, human speakers vary their pitch contour in subtle ways. AI voices in 2021 produced prosody that was consistent sentence-to-sentence in a way real humans never were. The eerily-uniform consistency was a tell.

4. Self-correction frequency. Real humans correct themselves several times per minute in any sustained speech — "the appointment is at — actually, let me check — 3:30". AI never did this. Removing self-corrections from human samples increased AI-detection rates immediately.

5. Conversational reciprocity timing. When two humans converse, the gap between one speaker finishing and the other beginning has a specific distribution — usually 200-500ms, occasionally with overlap. Synthetic voices in 2021 either responded too fast (instant) or too slow (1-3 second processing delay). Both broke listener expectation.

6. Regional accent fidelity. This was the most surprising finding. In smaller-market regions — particularly the American South — listeners detected AI faster when the voice's accent didn't match the local dialect, regardless of whether anything else was wrong. The brain's "this voice doesn't belong here" pattern was a faster detection trigger than any of the other five categories combined.

Why the Findings Made Existing AI Voice Tech Unsalvageable

By month nine of year one, we had a clear picture of what mattered. The next question was whether we could apply our findings to existing TTS systems — wrapping them in a "humanization layer" that added the missing imperfections after the synthetic voice was generated.

We tried. It didn't work. The reason has to do with how voice synthesis architectures from that era were built: they were optimized end-to-end for consistency. Adding imperfections after the fact produced output that sounded glitchy rather than natural, because the imperfections didn't match the underlying prosodic patterns the synthesis engine had baked in.

That conclusion ended year one and started year two. Building a voice engine from the ground up was a much bigger undertaking than anyone had anticipated — but it was the only path forward. The engineering implementation that followed is covered in Why Our AI Says "Umm".

The Founding Insight

If you compress the entire year of research down to one sentence, this is what you get:

Humans don't sound human because they're saying the right words. They sound human because they're saying them imperfectly, in patterns the listener has heard millions of times.

That sentence has shaped every product decision Futuro has made since. It's why the agents say "umm." It's why they breathe. It's why they self-correct mid-sentence. It's why we clone regional voices instead of using one neutral national accent. It's why the filler-word system inserts a 700-1500ms imperfection-laden audio clip the moment a caller stops talking, giving the reasoning system time to formulate a high-quality answer without producing the awkward silence that would give away the AI.

It's also why we built MasterMind the way we did — knowledge access during the imperfection, not before or after, because any pause for retrieval would announce the AI more loudly than any incorrect answer. And it's why every Futuro agent has access to 150+ tools — because a real human employee uses many tools naturally during a conversation, and an AI that only talked would feel robotic regardless of how good its voice sounded.

Year one's research is the foundation under all of it.

What This Research Enabled in Years 2-5

The research findings from year one became the design constraints for every subsequent year:

• Year two went into building voice-cloning technology that could reproduce a recorded human voice while preserving its imperfections — the opposite of every other voice-cloning effort at the time, which was trying to clean recordings up
• Year three went into building MasterMind so that knowledge retrieval could happen during latency-bridging filler clips rather than before responses started
• Year four went into deploying the architecture at scale across small businesses and refining the per-vertical knowledge patterns
• Year five (the current year) added the analytics platform, expanded the tool ecosystem to 150+ integrations, and culminated in the 94% indistinguishability double-blind study — the empirical confirmation that the year-one research had been correct

Without the year-one research, none of the rest would have worked. With it, every subsequent decision had a clear direction. You can book a demo and place a real call to a Futuro agent to hear what four years of building on year one's findings actually produces.

Frequently Asked Questions