AI voice comparison you can actually hear
Every voice model demo sounds great in the vendor reel. We give them the same script and judge what the reels hide: prosody under long sentences, emotional control, hard pronunciations, and latency you would ship with.
Text-to-speech crossed the "sounds human" bar for short, friendly sentences a while ago. What still separates models is everything real products throw at them: a 40-word sentence with two subclauses, a phone number followed by a Polish surname, sarcasm that should not sound cheerful, and a user who interrupts mid-reply.
We run voice the way we run everything here: one script, every model, full audio side by side. No cherry-picked lines, no vendor demo scripts — the exact input is published, and you hear complete takes, not highlight cuts.
This page is the guide: the dimensions we listen for, how the comparison works, and the roster we plan to run. Judging audio blind is even more important than judging text blind — brand expectations color what people hear.
What we listen for
Not "does a single sentence sound human" but whether stress, pacing and intonation survive long paragraphs. The tell is sentence number twelve: many models drift into a looping melody or robotic evenness the demo never shows.
Can the model actually deliver the emotion the text calls for — and can you steer it? We test explicit direction (whisper, urgency, deadpan) and implicit cues like sarcasm and bad news, where a default-cheerful voice fails instantly.
Names, brands, numbers, dates, units, acronyms, code-switching into foreign words — the unglamorous 10% that breaks products. A voice that nails "Grzegorz Brzęczyszczykiewicz called at 15:40 about the SQL query" earns its place.
Time-to-first-audio decides whether a voice agent feels alive or like a walkie-talkie. We measure first-byte latency and whether quality degrades in streaming mode, because conversational products live and die on this number.
Given a short reference sample, how close is the clone — and does it stay close across emotions and languages, or collapse into a generic voice with the right timbre? We also note each vendor's consent and safety gates, because cloning without them is a liability.
For realtime voice-to-voice models: interruption handling, backchannels, staying in character over minutes, and graceful recovery when the user talks over the model. This is where "TTS with a microphone" and true speech-native models separate.
How we run it
The shared rules — one prompt, one shot, blind votes — live at /method. Below is what is specific to this arena.
- 1One script, every model
A fixed script battery goes to every model unchanged: long-form narration, dialogue with emotional turns, a pronunciation gauntlet, and a realtime conversation scenario. The exact scripts are published with the results.
- 2Full takes, not highlight cuts
You hear the complete generated audio, including the awkward middle of long paragraphs where models drift. Latency numbers come from the same runs, measured, not quoted from pricing pages.
- 3Blind first, then vote
Audio plays without vendor labels, in shuffled order. You pick the voice you would ship, labels come off, and votes roll into a public tally — same mechanics as the coding and writing arenas.
Voice models on the test roster
What each family is known for, and where we expect the audio to be tested hardest.
| Model | Known for | What we watch | Status |
|---|---|---|---|
| ElevenLabs | Expressive TTS and cloning, the default benchmark | Whether control holds on long-form and non-English text | On the test roster |
| OpenAI | Realtime speech-to-speech and cheap capable TTS | Conversational robustness vs raw voice quality | On the test roster |
| Multilingual reach and native-audio Gemini voices | Prosody consistency across languages | On the test roster | |
| Cartesia | Ultra-low-latency streaming TTS | Whether speed costs naturalness | On the test roster |
| Hume | Emotion-first voice with fine-grained delivery control | Does measured emotion beat prompted emotion | On the test roster |
| Open weights | Kokoro, Chatterbox and friends — self-hostable voices | How close free gets to the paid tier | On the test roster |
This is the evaluation roster and what we plan to listen for, not scored results. Nothing here is ranked yet; the live arena is where the voices earn their place.
Voice model FAQ
▸How do you compare AI voice generators fairly?
Same script, every model, full takes side by side, judged blind before vendor labels show. Audio is judged even more by brand expectation than text, so hiding the label until after the vote is the core of the method.
▸Which AI voice model sounds most natural?
On short demo sentences, nearly all of them. The differences appear on long paragraphs, emotional turns, and hard pronunciations — which is exactly what our script battery tests. The arena results, not this guide, will answer the ranking question.
▸Why does latency matter so much for voice?
A voice agent that answers in 300 milliseconds feels present; the same voice at two seconds feels like a phone menu. Time-to-first-audio is the single number that decides whether realtime products are viable, so we measure it on every run.
▸Do you test voice cloning?
Yes — clone fidelity from a short reference sample, tested across emotions and languages, alongside each vendor's consent requirements. We only clone voices we have explicit permission for, and we treat missing safety gates as a negative finding.
▸Will you test speech-to-text too?
Transcription is a separate battery planned for the same arena: accuracy under noise, accents, code-switching and technical vocabulary. TTS and realtime voice-to-voice come first because that is where buyers have the least trustworthy information.
▸When does the voice arena launch?
After images and video. Voice needs an audio player UI and latency instrumentation the current arenas do not have, so the guide ships first and the side-by-side follows.
The method is already live
You can see the exact approach running today in the coding and writing arenas: one prompt, every model, real outputs judged blind. Voice gets the same treatment.
Open the coding arena