Which AI Receptionists Sound the Most Human? The Honest Answer Is a Test, Not a Name

Disclosure: Futuro builds VoiceAlive, an AI receptionist evaluated in this article under the same criteria and scripts as every other platform — with its limits stated in the evidence section. We never accept paid placement in any article; see our Publishing Principles. Every external claim was verified against the linked source on August 12, 2026.

Can any company honestly claim the “most human-sounding” AI receptionist?

No — and you should be suspicious of any vendor that does. There is no objective, industry-accepted scoreboard for “human-sounding.” The closest thing the speech sciences have is the Mean Opinion Score, a 1-to-5 listener-rating scale standardized in ITU-T P.800 and adapted for crowdsourced panels in ITU-T P.808. MOS is real, rigorous, and used by every serious speech lab on earth — and it is also a measure of preference, which means it moves with who is listening. The 2022 VoiceMOS Challenge showed that even state-of-the-art systems built to predict MOS scores degrade noticeably when the speakers or the listeners are ones the model has never seen. Rater panels disagree. Accents, age, hearing, and cultural expectations all shift the score. That is not a flaw in MOS; it is the honest nature of the question.

The lab results have also outrun the phone line. In 2024, Microsoft’s VALL-E 2 reported human-parity zero-shot text-to-speech on benchmark recordings — read that carefully: parity on recorded benchmark clips, in a lab, with no caller interrupting, no bad cell connection, no crying in the background, and no one asking a question the system wasn’t designed for. What reaches your customers on a Tuesday afternoon is not a benchmark clip. It is a live, bidirectional phone call where human conversation research shows people expect a response in roughly a quarter of a second and silently judge anything slower.

So this article does something different from the usual “top 10 most realistic AI voices” listicle. Instead of telling you which platform sounds most human — a verdict we cannot honestly give and neither can anyone else — we are handing you the test we use ourselves: seven measurable criteria, five scripted test calls, and a scoring sheet you can run against any platform, including ours, in about 45 minutes. Every criterion is defined by a failure you can hear with your own ears, not an adjective you have to take on faith. For the broad category comparison across features, pricing, and use cases, see our buyer’s guide to the best AI voice agent platforms by use case; this article deliberately stays on the single question of human-likeness.

What seven criteria actually measure “human-sounding”?

Strip away the marketing and “sounds human” decomposes into seven things you can hear, time, and count. Each criterion below has a definition, a failure signature — what it sounds like when a platform gets it wrong — and the research or standard it rests on. None of them require lab equipment; a stopwatch app and a careful ear are enough for the version in this article.

1. How fast does it answer? (Response latency)

What it is: voice-to-voice latency — the gap between the caller finishing a sentence and the agent starting its reply, including turn detection, speech recognition, the language model, synthesis, and the phone network. Stivers et al.’s cross-language study in PNAS found humans across ten languages take their next turn after a median gap of roughly 250 milliseconds. Levinson’s work on conversational timing explains why: listeners begin planning their response while the other person is still talking, so a system that waits for silence and then starts processing is already late by human standards.

What failure sounds like: a walkie-talkie rhythm. You stop talking; one-Mississippi, two-Mississippi; then the reply. Callers start saying “hello? are you still there?” or talking over the agent, which makes the next exchange worse.

What the numbers say: engineering analyses put a well-tuned 2026 stack near 600 milliseconds end-to-end and a mediocre one at 1.2–1.8 seconds (Retell’s stack breakdown); Hamming AI’s component guide sets the production target under 800 ms and notes the language model alone is typically 70% of the budget. The most useful public comparison is Telnyx’s 2026 claims-versus-measured roundup: it places vendor-published figures next to a third-party test of 500 production calls per platform — Vapi targets p50 under 500 ms but measured 720 ms median, Retell claims ~600 ms and measured 680 ms median (the lowest in that test), and Bland’s 400 ms homepage figure measured 850 ms. Two lessons: vendor numbers are real but best-case, and even the best measured medians sit at roughly triple the human norm. Master of Code’s pipeline analysis adds the structural point: sequential architectures stack delays to 800–2,000 ms before telephony, and one multi-million-call analysis puts the industry median for cascaded systems at 1.5–1.7 seconds.

2. Does the voice carry melody? (Prosody and intonation)

What it is: pitch movement, rhythm, stress, and pacing — the difference between a sentence that is said and one that is read. Prosody carries meaning: the same words can reassure, rush, or brush off a caller depending on the melody.

What failure sounds like: flat, even delivery with the emphasis landing on the wrong syllable of the wrong word; or the opposite — exaggerated radio-announcer bounce on a sentence about a billing error. There is also a subtler failure: a voice that is 95% right. Mori’s uncanny-valley research documented that near-human-but-not-quite agents can provoke more discomfort than obviously synthetic ones, because the ear keeps catching the seams.

How to judge it: this is the criterion MOS was built for — P.800 formalized the 1-to-5 listener scale, and P.863 (POLQA) defined the objective algorithmic version used to grade telephony speech. McAleer, Todorov, and Belin’s voice-perception work adds the uncomfortable detail that listeners form personality judgments from a voice within a few hundred milliseconds of hearing it — before a single word of substance lands. For the engineering view of how one platform approaches natural rhythm — including deliberate breaths and disfluencies — we published our own approach in why our AI voice breathes, stutters, and sounds human.

3. Can you interrupt it? (Barge-in handling)

What it is: whether the agent notices when you talk over it, stops, and responds to your interruption — instead of plowing to the end of its paragraph. Human turn-taking, as Sacks, Schegloff, and Jefferson’s foundational 1974 paper described it, runs on minimal gap and minimal overlap; people negotiate the floor constantly, and interruption is normal, not exceptional.

What failure sounds like: you say “wait, actually—” and the agent keeps reading for four more seconds, then asks a question you already answered. Worse variants: the agent stops but restarts its whole script from the top, or stops and goes silent because your interruption confused its state machine.

Why it is hard: Skantze’s 2021 review of turn-taking in conversational systems catalogs the trade-off — detect end-of-speech too eagerly and the agent cuts callers off mid-word; too conservatively and you get the walkie-talkie problem from criterion 1. Recent research on predictive turn-taking found listeners prefer systems that anticipate turn endings from intonation and syntax over systems that simply wait out a fixed silence threshold — the same thing human listeners do.

4. Does it say names and trade words correctly? (Pronunciation)

What it is: correct, confident first-attempt delivery of personal names (“Siobhan,” “Nguyen,” “Xochitl”), your business name, street addresses, and the vocabulary of your trade — “balayage,” “occlusion,” “escrow,” “bevel gear.”

What failure sounds like: the agent mangles the caller’s name in the greeting and never recovers; or it spells an unusual word out letter by letter, which is honest but immediately announces “machine.” This is the criterion where the uncanny valley bites hardest — the more human the voice, the more jarring a robotic mispronunciation feels by contrast.

Why it matters commercially: a receptionist’s first job is recognition. Getting a returning customer’s name right is the cheapest trust signal in the entire call, and getting it wrong is the fastest way to make a caller feel processed rather than served. When you run the test calls below, plant one genuinely difficult name and one trade-specific term, and keep score across platforms.

5. Does it remember what you said thirty seconds ago? (Context retention)

What it is: within-call memory. The caller gives their name once, their appointment date once, a constraint once (“anything after 3pm”) — a human receptionist weaves all three into the rest of the conversation. So should the machine.

What failure sounds like: “Can I get your name?” — asked twice in one call. Or the caller says “I’m moving my Thursday appointment,” confirms Tuesday at 4, and the agent closes with “so that’s Thursday at 4.” Small slips, enormous signal: the caller now double-checks everything else the agent says.

How to test it: plant three facts early in the call, change one of them midway, and count how many the agent uses correctly at the end. Our first scripted test call does exactly this, and it is deliberately the easiest of the five — most platforms pass it, which is what makes the ones that don’t so instructive.

6. What happens when it doesn’t understand? (Recovery from ambiguity)

What it is: the behavior in the gap between “understood perfectly” and “total failure.” Real callers mumble, ask two things at once, use slang, or ask something the business never anticipated. Turn-taking research shows humans handle this with short, specific repair questions — “did you mean this Friday or next?” — and a good agent does the same.

What failure sounds like: the confident guess (the agent books the wrong thing with total assurance), the generic stall (“I’m sorry, I didn’t get that” on a loop), or the non-sequitur — answering a question you didn’t ask because it pattern-matched the nearest script line. The confident guess is the dangerous one: the other two failures are annoying, but a confidently wrong booking creates a real-world mess a human has to clean up.

How to test it: ask one deliberately unusual, two-part question — our third test call is “do you take cash, and is the entrance step-free?” — and listen for whether the agent answers both halves or quietly picks the easy one.

7. Does it know when to get a human? (Safe escalation)

What it is: recognizing the edge of its competence — an upset caller, a sensitive situation, a request outside its authority — and handing off to a person with the context of the call intact, rather than trapping the caller in a loop.

What failure sounds like: the caller says “this is the third time I’m calling about this” and the agent chirps a scripted apology and re-offers the same FAQ. Crolic et al.’s 2022 Journal of Marketing study found that highly human-like bots can actually backfire with already-angry customers — the more the bot sounds like a person, the more the caller expects human-level understanding, and the angrier they get when it isn’t there. The design implication is uncomfortable but clear: sounding human raises the bar for knowing when to hand off to one.

How to test it: ask for a manager — politely the first time, firmly the second. Our fifth test call scripts both attempts, because many platforms handle the polite version and fail the firm one.

How should you weight the seven criteria?

There is no universal weighting, and we won’t pretend one exists. For a first-pass screen we use latency, prosody, and barge-in at 20% each — they are the three a caller perceives within the first ten seconds — context retention and recovery at 15% each, and pronunciation and escalation at 5% each. That is a screening rubric, not a verdict rubric: if your calls are emotionally charged (medical offices, funeral homes, crisis lines), escalation deserves far more than 5%, and if your clientele skews multilingual, pronunciation climbs the list. The “best for whom” matrix below re-weights by buyer profile, and the test-call section shows how to apply the rubric without changing a single script.

What are the five test calls we run on every platform?

These five scripts are the heart of the framework. They are deliberately business-agnostic — they work against a dentist’s demo line, a salon’s, a plumber’s, or any vendor’s published demo number — and each one is engineered to stress a specific criterion. Run all five, word for word, against every platform you are evaluating, on the same day, over the same kind of phone connection. For a vertical application of the same idea, our hair-salon comparison runs five salon-specific calls through the same logic.

Test conditions (hold these constant): call from a normal mobile phone, not a landline with perfect audio; call each platform within the same two-hour window so network conditions are comparable; do not warn the agent you are testing; and have a second person listen on speaker where possible, because a single rater is a single opinion. Methodology finalized August 2026; any future revision will be logged in the changelog.

Test call 1: Can it handle an appointment change with a moving constraint?

Script (read naturally, not robotically): “Hi, I need to move my appointment this Thursday — it’s under Maria Okonkwo — to sometime next week, but it has to be after 3pm because of my work schedule. Actually, wait — my coworker just texted, could we also see if Tuesday morning works instead?”

What it stresses: context retention (criterion 5) above all — the caller plants a name, a date, and a constraint, then revises the constraint mid-call. It also touches latency and recovery.

A pass sounds like: the agent tracks all three facts, acknowledges the revision without confusion (“no problem, let’s look at Tuesday morning instead”), and closes with an accurate readback: name, new day, correct time window.

A fail sounds like: being re-asked for the name; the “after 3pm” constraint silently surviving into the Tuesday-morning option where it makes no sense; or a closing readback that mixes Thursday and Tuesday.

Test call 2: What happens when you interrupt mid-sentence?

Script: call the demo line and let the agent begin its greeting or its first long answer. Two seconds into its sentence, cut in with: “Sorry — quickly, is there parking?” Then stop and listen.

What it stresses: barge-in handling (criterion 3) — the single most physical test in the set, and the one predictive turn-taking research says separates systems built on anticipation from systems built on silence timers.

A pass sounds like: the agent stops within roughly a syllable or two of you starting, answers the parking question, and then gracefully offers to continue or pick up where it left off. Bonus marks if it doesn’t restart its previous sentence from the top.

A fail sounds like: it talks over you for three or four more seconds — the two-voices-at-once collision every caller recognizes instantly — or it stops but then behaves as if the parking question never happened. Skantze’s review documents both failure modes across two decades of spoken-dialogue systems; they remain the norm, not the exception.

Test call 3: Can it answer a question it wasn’t expecting?

Script: “Two quick things before I book — do you take cash, and is the entrance step-free? My mother uses a walker.”

What it stresses: recovery from ambiguity (criterion 6), with a pronunciation and prosody rider. The question is mundane for a human receptionist and surprisingly hard for an agent: it’s two questions in one sentence, at least one of which (“step-free”) is rarely in any FAQ, and it carries an emotional context that a good voice should acknowledge.

A pass sounds like: both halves addressed — a straight answer on payment, and either a real answer on accessibility or an honest “I’m not certain about the entrance, let me note that for the team and have someone confirm” — delivered with a beat of warmth for the mother, not a flat pivot. The honest-deferral version is a pass, not a fail; guessing an answer about wheelchair access would be the worst possible outcome.

A fail sounds like: answering only the cash half; answering a different question entirely; or confidently inventing an accessibility fact. Rate the confident invention as a zero on recovery — this is the failure that puts a caller with a walker at the bottom of a staircase.

Test call 4: How does it treat an emotional caller?

Script (deliver with real hesitation in your voice): “Hi… I need to cancel my dad’s appointment on Friday. He’s — he’s actually in the hospital, so I’m not sure when we’ll rebook. Sorry, I’m a bit scattered.”

What it stresses: prosody and escalation together (criteria 2 and 7). There is no correct information to retrieve here; the test is whether the voice changes — slower, softer, unhurried — and whether the agent completes the simple task (cancel Friday) without making the caller repeat anything or pushing a rebooking script.

A pass sounds like: an acknowledgment that is brief and human-scaled (“I’m sorry to hear that — I’ll take care of the cancellation right now”), the task done, and a gentle, non-salesy close — optionally flagging a human follow-up. Crolic et al.’s findings are directly relevant: with emotionally charged callers, a bot that sounds warmly human but fails to respond humanly makes the interaction worse, not better.

A fail sounds like: cheerful default energy (“Great! I can help with that!”), a rebooking pitch to a man whose father is hospitalized, or — the classic — “Is there anything else I can help you with today?” read in the standard sign-off cadence.

Test call 5: Does asking for a human actually reach one?

Script: first, politely: “This is a bit complicated — could I talk to a person?” Whatever happens next, follow once, firmly: “I understand, but I’d really like to speak with a manager.”

What it stresses: safe escalation (criterion 7) — the criterion that protects your business when everything else fails. Note that “pass” does not require an instant live transfer; many excellent setups take a message and guarantee a callback window. What matters is that escalation is real.

A pass sounds like: the agent respects the first or second request, explains exactly what happens next (“I’m flagging this for Sarah, who manages the front desk — she’ll call you back within the hour at this number”), captures your details without making you repeat the call’s history, and confirms the handoff in a closing summary or text.

A fail sounds like: a loop (“I can help with that!” repeated with rising customer blood pressure), a dead end (“our team will reach out” with no name, no window, no confirmation), or — most damning — the agent arguing with the request. If you want a reference point for what the alternative looks like, Smith.ai and Ruby sell human receptionists with AI assistance rather than AI with human backup — a structurally different category worth knowing exists before you score this criterion.

Why five calls and not one: any single call can flatter or sink a platform by luck. Five calls through seven criteria produce 35 scored observations per platform — enough that the pattern, not the anecdote, is what you end up buying. And because the scripts are fixed, your results are comparable against ours, against a colleague’s, and against your own re-test six months from now.

How do you score the calls without fooling yourself?

A framework you can game isn’t a framework. Two habits make this one honest: fixed anchors for each score, and enough structure that your expectations — especially if you’re already leaning toward a platform — can’t quietly rewrite what you heard.

What does a 0, 1, or 2 actually mean?

Score every criterion on every call as 0, 1, or 2 — never a 1.5, never a “2 minus.” The anchors:

• 2 — indistinguishable behavior. A caller with no reason to suspect AI would not have clocked it on that criterion. For latency, replies land inside roughly one second; for barge-in, the agent yields within a syllable or two; for recovery, it asks a short, specific clarifying question.
• 1 — noticeable but tolerable. You noticed the machine, but the task completed and a patient caller would stay on the line. A two-second pause before each answer. A barge-in that eventually registers. A recovery that takes two attempts.
• 0 — task-breaking or trust-breaking. The failure a caller remembers and repeats. The loop, the confident wrong answer, the talked-over interruption, the mangled name in a condolence call.

Thirty-five observations (5 calls × 7 criteria) produce a maximum of 70 points. We treat 56+ (80%) as “deployable for most businesses,” 42–55 as “deployable for structured, low-emotion call types,” and below 42 as “keep looking.” These thresholds are our editorial judgment, not an industry standard — ITU-T P.808 and the UTMOS line of work show how far more rigorous subjective testing can go — but they are fixed, published, and applied to ourselves, which is the part that matters for comparability.

What are the rater rules that keep it fair?

• Score during the call, not after. Memory edits toward the impression you wanted. Tick the sheet in real time.
• Two raters where possible, and they don’t confer. Independent scores, averaged at the end. Where raters split by two full points, re-run that call — the disagreement usually means the script wasn’t delivered identically.
• Same day, same phone, same hour window. Latency in particular moves with network conditions; a platform tested at 9am on office fiber and another at 6pm on a congested cell tower aren’t being tested, they’re being weather-reported. Pipeline analyses show PSTN overhead alone adds hundreds of milliseconds that no vendor controls.
• Never score a platform you can’t identify. Record the vendor, the product tier, the date, and the demo line used on the sheet itself. An unattributed observation is worthless to the next person.
• Keep the recordings where lawful. Consent rules for recording calls vary by jurisdiction — one-party versus all-party consent — and they apply to test calls too. Check your state’s rule before you hit record, or take contemporaneous notes instead. This is also why our own published results (below) currently ship as observations rather than audio.

The scoring sheet itself — the five scripts, the anchors, the observation grid — is published in this article under a CC-BY-4.0 license and marked up as a machine-readable dataset, precisely so that other evaluators can run identical conditions and publish contradicting results. We would rather be checked than believed.

What can we honestly publish today — and what can’t we?

This is the section most “comparison” articles fake. They present tables of impressionistic adjectives — “Bland: robotic,” “Vapi: natural” — with no scripts, no conditions, no dates, and no way for you to check a word of it. We won’t do that. Here is exactly what evidence exists today, August 2026, and exactly where its edges are.

What we publish: the framework, and our own front door

Published today, in full: the seven criteria, the five verbatim scripts, the scoring anchors, the conditions, and the rater rules — everything in this article, licensed CC-BY-4.0 so anyone can reuse it. Also public: our own demo line. Futuro’s VoiceAlive agent — our product, and we will always say so — answers live calls on our public demo and through the 7-day trial, which means you can run all five test calls against us today, score us with our own sheet, and publish the result anywhere you like. We would rather be measured harshly and publicly than praised vaguely.

For the record on what we claim in this article and nothing more: our 1,000-participant double-blind listening study documents listener response to VoiceAlive’s voice — that study’s methodology lives at the link and is deliberately not restated here. It says nothing about any competitor. Claims of category-wide superiority would need exactly the cross-platform data we are telling you we don’t yet have, and we don’t make them.

What we withhold, and why: recorded competitor calls

We ran draft versions of these five calls against several platforms’ public demo lines in July–August 2026. We are not publishing the recordings, the platform-by-platform observation table, or any scores from those calls yet, for two reasons that have nothing to do with being polite to competitors:

• Recording-consent law. Publishing audio of a call — even to a public demo line — implicates state wiretap and consent rules that vary by jurisdiction, and a competitor’s terms of service may separately restrict recording and republication of their demo. Our counsel’s review of what we can publish, in what form, was still open on August 12, 2026. Observations-as-text are our likely landing zone; raw audio needs the legal answer first.
• Sample size. Five calls against a demo line whose configuration we don’t control is a screen, not a verdict. Publishing scores from n=1 configurations, however honestly labeled, would hand readers a leaderboard — and this article’s entire premise is that the honest deliverable is the test, not the leaderboard. The best third-party latency work we cite makes the same point about its own 500-call study: directionally comparable within the test, not a universal ranking.

When the consent review closes and the quarterly re-run (see governance) produces a properly-sized, dated observation set, we will publish the table here with a changelog entry. Until then: any site — including ours — that shows you a definitive “human-ness ranking” of platforms without scripts, dates, and conditions is showing you an opinion wearing a lab coat.

Where we weren’t the best — said out loud

Here is the uncomfortable sentence this article owes you. In that same July–August draft screening, and in the public benchmark data anyone can read, Futuro does not publish independent, third-party latency or interruption measurements of our own agent — and on the criterion of published, verifiable performance data, platforms like Retell (which publishes its stack budget and has been third-party measured at 680 ms median in the Tested Media study cited above) are ahead of us. Our listening study measures how our voice is perceived; it is not a latency benchmark and we won’t dress it up as one.

What we do about it is also on the record: our demo line is public, our trial is open, and this article hands you the stopwatch. If VoiceAlive’s barge-in handling stumbles when you run test call 2, you will hear it, you can score it 0, and you should. Commissioning an independent, methodology-disclosed latency and interruption study of our own stack — the same standard we cite competitors’ numbers from — is on our governance roadmap below, and the changelog will carry the date it lands. Evidence, not adjectives, is the whole game.

So who is each type of platform actually best for?

Because we publish no cross-platform ranking, the honest version of “which should I buy” is a matching exercise: your call profile determines which criteria dominate, and the criteria tell you what to test hardest. Six profiles cover most buyers. Full platform-by-platform evaluations — pricing, integrations, trade-offs — live in our buyer’s guide by use case and customer-service receptionist guide; this matrix deliberately re-derives everything through the human-likeness lens and points there for the rest.

Where does Futuro actually fit?

In the profiles we actually win, we’ll say so plainly: businesses whose calls carry emotional weight and whose owners want the agent modeled on their best human receptionist — the Human Staff Mirroring approach behind VoiceAlive (our product). The profiles we don’t win are equally plain: if your top criterion is a published, third-party-verified latency floor, the measured data in the Telnyx roundup currently favors others, and we said so in the evidence section. If you want a build-it-yourself stack, you want Vapi or Retell, not us.

What this matrix deliberately is not

It is not a ranking, a star table, or a set of recommendations — it is a mapping from buyer profiles to the criteria that should dominate your scoring sheet. Every “evidence supports” cell either cites measured data with its scope labeled, or names the structural trade-off (human-first pricing, no-code configuration risk, tuning burden) that no vendor disputes. Where a platform appears by name, it appears because a cited third-party measurement or an architectural fact requires the name — never as an endorsement or a takedown.

Who is not in this matrix, and why

Enterprise contact-center platforms (PolyAI-class), carrier infrastructure vendors, and pure TTS engines like ElevenLabs’ stack are out of scope: they sell to different buyers at different layers of the stack, and including them would blur the receptionist question this article exists to answer. Deepgram’s low-latency guide explains the component layer well if you’re building rather than buying. Vertical-specific picks — salons, trades, medical — live in the buyer’s guides linked above rather than here, because the rubric that matters changes by industry.

How do you run this test yourself in 45 minutes?

Everything above is designed to be executed by a busy owner between appointments, not by a lab. Here is the protocol as a single sitting.

What are the exact steps?

1. Pick three demo lines (5 minutes). Any three platforms whose demos answer a real phone call. Include ours — futurocorp.com/demo or the 7-day trial — because a framework the seller won’t submit to is marketing, and this one isn’t. Add one platform from whichever profile row in the matrix matches you, and one wildcard you found on your own.
2. Print or copy the scoring sheet (5 minutes). Five calls across seven criteria, 0/1/2 anchors, one row per observation. Note the date, your phone, and each platform’s name and demo source at the top.
3. Run the five scripts per platform, in order, back to back (25 minutes). Word for word — the scripts only work if they’re identical. Score during the call. Don’t re-run a flubbed delivery; note it and move on, because a flubbed delivery happens to real customers too.
4. Tally, then re-run exactly one call (10 minutes). Total each platform out of 70. Then re-run whichever single call most surprised you — good or bad. One surprise is an anecdote; the same surprise twice is a pattern.

That’s it. You now hold more real evidence about those three platforms than any published ranking can give you, because it was measured on your ear, your phone, and the questions your customers actually ask.

What fairness rules apply when you publish your results?

If you write up your test — and we hope people do — four rules keep it credible. Date every observation; demo lines change monthly. Name the exact demo source (a public number, a trial account, a sales-engineered demo) because they are not the same product. Publish your scoresheet, not just your conclusions, so others can re-run your calls. And apply your rubric to your own current phone setup too — the honest control group is whatever answers your line today. If your write-up contradicts this article’s expectations on any platform including ours, that is the framework working as designed. Voice preference is subjective; these results are a repeatable evaluation framework, not a universal ranking — yours included.

How do we keep this framework honest over time?

A methodology that isn’t maintained becomes stale marketing within a year. Three commitments keep this one alive, and they’re written here so readers can hold us to them.

Who owns the methodology?

Brandon Gillespie, Futuro’s founder (bio), owns this protocol editorially, with review by the Futuro editorial team. Ownership means: the scripts don’t change silently, the anchors don’t drift, and any dispute about what a score means gets resolved against the published anchors rather than redefined after the fact. If a criterion needs to change — because platforms genuinely fix barge-in and the differentiator moves elsewhere, say — the change happens through the changelog, not through a quiet edit.

When does the evaluation re-run?

Quarterly. The next full re-run of the five calls against the rotating platform set is scheduled for November 2026, then February, May, and August 2027. Each re-run applies identical scripts and conditions, adds any platform that has become significant to our readers, and — once the recording-consent review described in the evidence section closes — publishes the dated observation table here. The same quarterly cycle also re-verifies every external claim this article relies on: the third-party latency figures, the academic citations, and the live status of every demo line we reference.

What has changed so far?

Changelog:

• August 12, 2026 — Initial publication. Seven criteria, five test calls, scoring anchors, and rater rules finalized. Competitor audio and observation table withheld pending recording-consent legal review (status noted in the evidence section). All external benchmark figures re-verified against source documents on this date.

Future entries will be appended, dated, and never silently edited — a changelog that can be revised is a changelog that means nothing.