Picking a Synthetic Voice: Five Attributes That Beat Natural

Ask someone auditioning AI voices what they're listening for and almost everyone says the same word: natural. It's not a bad instinct, but it's a bar nearly every current TTS model clears on a single ten-word sample line. The problem is that a ten-word sample line is not a script. Voices that sound identically convincing on "the quick brown fox jumps over the lazy dog" separate hard once you put them through an actual 90-second narration with a pace change, a parenthetical aside, an emotional beat, and a sibilant-heavy product name repeated four times. Natural is table stakes. Here are five attributes that actually predict whether a voice survives real use.

1. Pace elasticity

A script isn't read at one speed. A hook needs urgency, a list needs a beat between items, a CTA needs to slow down so it lands. Pace elasticity is whether a voice can move between those speeds within one take without the delivery breaking character — versus a voice that has one comfortable tempo and starts to sound mechanical the moment you push it faster or slower than that.

This is a real, exposed control on several engines rather than something you infer from a sample: Cartesia's Sonic models list "Speed Control" as a named feature directly alongside "Emotion Control" in Versely's own model catalog. That's worth auditioning deliberately rather than trusting on the strength of a single-tempo demo — generate the same script at three speed settings and listen for where the voice starts to lose its own rhythm rather than genuinely accelerating.

2. Emotional range

Versely's own glossary entry on voice stability gets at the mechanic behind this directly: stability is "the control that decides how much a synthetic voice varies its delivery — steady and predictable at one end, expressive and unpredictable at the other." Every provider names this control differently — stability on one engine, exaggeration on another, expressiveness on a third — and critically, some invert the direction, so a numeric value that means "calm" on one model can mean "wild" on another. The attribute worth testing isn't whether a voice CAN sound excited on command. It's whether it can move credibly between register — flat delivery, a mild aside, a genuine emotional peak — inside the same script, without every setting above the midpoint collapsing into the same generic "energetic" read.

3. Breath behaviour

Real speech isn't a continuous waveform of words — it has breath, pause, and the small non-verbal sounds that sit between clauses, and their absence is one of the fastest tells that something is synthetic even when the words themselves are delivered perfectly. Versely's audio tags glossary entry covers the mechanism some engines expose for this directly: bracketed or angle-bracketed markers written into the script — a laugh, a whisper, a pause — that tell the model how to deliver the words immediately around them, positioned exactly where the change should happen rather than applied to the whole take. The catch is that this vocabulary is entirely provider-specific, with no shared standard across engines, and a model that doesn't support tags will simply read the brackets aloud as text. Test breath behaviour specifically at a clause boundary — a comma, a dash, a parenthetical — because that's where its absence or presence is most audible, not in the middle of a clean declarative sentence.

4. Consonant clarity

This is the attribute most sample-line auditions never catch, because most sample lines are chosen for how they sound at a comfortable pace, not for how they hold up under pressure. Push the pace up and listen specifically to sibilants (s, sh sounds) and plosives (p, b, t) — the consonants that carry the most information at speed and are the first thing to blur or clip when a model is straining. A voice that's warm and clear at a relaxed narration pace can turn genuinely hard to parse at 1.3x, which matters enormously for any script with a brand name, a SKU, or a spelled-out URL in it. Run the exact phrase your script actually needs said clearly — the product name, not a generic pangram — before committing to a voice for anything with specific vocabulary in it.

5. Long-read consistency

The failure mode here isn't a bad line, it's drift: a voice that opens a two-minute narration in one register and has quietly shifted pace, tone or energy by the closing line, without any single sentence sounding wrong in isolation. It's the reason the voice-stability glossary entry specifically recommends high stability "for narration and anything long," reserving the lower, more expressive settings for character work and short punchy lines where a little unpredictability is a feature rather than a defect. Long-read consistency is also a selection problem, not just a settings problem: it's easy to end up mid-script having picked slightly the wrong voice from a very large shortlist, which is a real risk given how big these rosters actually are.

The rosters are bigger than most people account for

Versely's own voice surface, per an app-data snapshot taken 2026-08-11, spans 22 audio models across seven-plus providers — not every one a narration voice in the traditional sense (a couple are sound-effect or voice-conversion tools), but the speech engines proper cover MiniMax, ElevenLabs, Gemini, Grok, Cartesia, Inworld and Qwen. MiniMax alone contributes 148 uniquely named voices spread across 33 style buckets. That's not a shortlist you audition exhaustively — it's a field you need to narrow by hard constraint before you narrow it by taste.

Language is the fastest real filter, and coverage varies far more than most people expect: ElevenLabs covers 30 languages, Gemini 24, Grok 16, Cartesia 15, Inworld 15, and Qwen 10, per the same snapshot. If the brief is a Spanish-language narration, that constraint alone removes most of a 148-voice roster before a single sample gets played — filter by language coverage first, and only audition the five attributes above within whatever's left.

Latency is a separate axis — don't conflate it with quality

A handful of models in the catalog are explicitly built and named for speed rather than for winning a blind quality comparison: cartesia-sonic-3-5 and cartesia-sonic-3 are Cartesia's current top-ranked and prior-generation Sonic models, while chatterbox-tts-turbo and eleven-labs-speech-turbo are named, turbo-specific variants built for fast generation. That's a genuinely different axis from the five attributes above, and it's worth auditioning separately rather than assuming a "turbo" label tells you anything about where a model lands on pace elasticity or breath behaviour — it tells you about render speed, which matters enormously for a real-time or high-volume pipeline and not at all for a one-off hero narration recorded well ahead of a deadline.

Versely walkthrough: auditioning on all five at once

The efficient way to run this isn't five separate tests — it's one script built to expose all five attributes in under 90 seconds, run across a short, language-filtered list of voice_id candidates through Versely's generate_speech tool, which takes text, model, voice_id, emotion and style_instructions as first-class fields rather than a single flat prompt:

"Generate this 80-word script with three voices from [shortlist]: a flat opening line, a fast three-item list, a quiet aside in parentheses, one line with our product name repeated twice, and a slower closing line. Use emotion to mark the aside as warm and the opener as neutral."

Play the three results back to back on the exact same script rather than three different demo lines — that's the only way pace elasticity, emotional range, breath behaviour, consonant clarity and long-read consistency actually become comparable instead of anecdotal. Whichever voice survives the fast list without blurring the product name and hasn't drifted by the closing line is the one that'll hold up in the finished piece, not just in the pitch.

The takeaway

"Does it sound natural" stopped being a useful filter once most models cleared it. The five attributes that actually separate a usable voice from a merely convincing one only show up under the conditions a real script creates: a pace change, an emotional beat, a breath, a hard consonant at speed, and enough runtime for drift to become audible. Build the audition script to expose all five before you commit — see Versely's voice-over tooling and the current best text-to-speech model roundup for a starting shortlist, then run your own 80-word test rather than trusting someone else's ten-word one.