Expressiveness
Emotion and emphasis. Stressing the right words, sounding like it means what it says instead of reading text aloud.
The Humanness Index™
Which voice model sounds the most human?
Sounding human is hard to measure, but it's what decides whether a call works. We clone one voice onto every model and play them blind against a real human, so you can hear which ones pass.
Read the whitepaperWhich voice sounds more human?
Same voice, different models.
So I can see here that the package was marked as delivered on Tuesday, but if you're saying it never arrived then what we'll do is... let me just. Yeah, I'm going to open a lost package investigation for you. That usually takes about forty-eight hours to resolve.
vs
←→ play each side · space vote, then next pair
How it works
Step 1
Same voice, every model
We clone one conversational voice onto every model, so you're judging the model, not its demo reel.
Step 2
You listen blind
Two voices, same line, no labels. Pick the one that sounds more human.
Step 3
A real human sets the bar
Blind votes are fit into a rating, with a real human at 100. The higher the score, the more human the model sounds.
Humanness Rankings
Humanness distribution
21 Models11 providers13000 unique votes
Color = rank Average
Why latency matters. A voice that lags breaks the conversation, no matter how human it sounds.
| Likely Rank | Model | Listen | ||||||
|---|---|---|---|---|---|---|---|---|
| Baseline | Human | Homo Sapien | 100 | 1293 | — | — | 589 | |
| #1–6 | Speechify | Simba 3.2 | 97 | 1284 | 428 ms | $10 | 404 | |
| #1–4 | ElevenLabs | Eleven v3 | 97 | 1283 | 758 ms | $100 | 574 | |
| #1–7 | xAI | Grok TTS | 93 | 1270 | 460 ms | $15 | 559 | |
| #1–8 | MiniMax | Speech 2.8 | 92 | 1268 | 325 ms | $60 | 543 | |
| #2–8 | MiniMax | Speech 2 HD | 89 | 1259 | 357 ms | $100 | 524 | |
| #2–8 | Canopy Labs | Orpheus | 89 | 1258 | — | Open source | 529 | |
| #3–9 | Fish Audio | S2.1-Pro | 87 | 1253 | 141 ms | $15 | 396 | |
| #3–8 | xAI | Grok TTS (Streaming) | 87 | 1251 | 285 ms | $15 | 530 | |
| #8–12 | Inworld | TTS-1.5-max | 78 | 1224 | 337 ms | $35 | 453 | |
| #9–13 | ElevenLabs | Flash v2 | 76 | 1216 | 226 ms | $50 | 458 |
The Index only includes models that support voice cloning: each battle plays the same cloned source voice through both models, so the comparison is head to head and fair. The Index is an open benchmark and is independent of the Vapi product: a model appearing here does not mean it is available in Vapi, and availability in Vapi plays no part in scoring. Don't see your model on this list? Contact us at humannessindex@vapi.ai.
What we Listen for
What makes a voice sound human?
Humanness doesn't break down into features. You either believe there's a person on the other end, or you don't. When that belief breaks, it's usually because of one of these.
Tone & prosody
The intonation, rhythm, and melody of speech. The natural rise and fall of how people actually talk.
Artifacts
The little human sounds: breaths, stutters, natural pauses. A voice with none of them sounds too clean to be real.
Why trust this benchmark?
Any model can sound good on its own demo voice. The real test is how it handles your use case. We clone one voice across every model so the comparison is fair. Models that can't clone a voice can't be tested fairly, so they're not listed.
Most Human Models
#1 · Humanness leader
Speechify Simba 3.2 currently leads the Humanness Index™. Across blind listening tests it's judged the most human-sounding voice in the field, the kind of delivery that holds up with real callers in production, not just in a demo.
#2
Humanness97
- Latency
- 758 ms
- Languages
- 70+
- Votes
- 574
#3
Humanness93
- Latency
- 460 ms
- Languages
- 20
- Votes
- 559
Voting in progress
Rankings are provisional
We're keeping the podium under wraps while the votes come in. Listen and vote above, and the most human models reveal once the standings settle.
Why this exists
Picking a TTS model for a voice agent comes down to one thing: does it sound human enough that people forget they're talking to software? You can't get that from demos or vendor claims. So we made it measurable and took the call out of our own hands: one voice cloned onto every model, played blind with no names attached, scored against a real human by the people who hear it.
Find the most human-sounding voice for your agent.
Compare the models in blind tests, read the methodology, or get in touch.
Build a TTS model? Add yours to the Index.