Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Most natural TTS voices in 2026: three naturalness measures side by side
On Artificial Analysis listener preference, ElevenLabs Eleven v4 Turbo, Cartesia Sonic 3.6, and Inworld Realtime TTS-2 lead the ranked models, starting with the highest Elo. Speko and Cekura measure naturalness differently and are shown beside it.
The three scales differ: Elo comes from blind listening by people, Speko from blind A/B votes on phone-agent lines, and Cekura from a listener model's estimate. They are never added or averaged, and a model measured by one source only is not penalised for the others.
How the rows are ordered. Ordered by Artificial Analysis listener Elo, highest first; models without an Elo come last, in ranking order. Two Elo values whose 95% intervals overlap are statistically tied.
| Model | Listener Elo Artificial Analysis, with 95% interval | Customer-service Elo Artificial Analysis, customer-service category | Naturalness (Speko) Blind A/B votes | Naturalness (Cekura) Estimated by a listener model, 1 to 5 | Rank | Speed tier From the Coval P95 |
|---|---|---|---|---|---|---|
 ElevenLabs Eleven v4 Turbo | 1,333 ±16 | 1,372 ±33 | 1,670 | 4.42 | 8 | Fast enough |
 Cartesia Sonic 3.6 | 1,292 ±14 | 1,331 ±28 | 1,578 | 4.41 | 14 | Visible delay |
 Inworld Realtime TTS-2 | 1,267 ±14 | 1,311 ±28 | 1,561 | 4.52 | 6 | Fast enough |
 Inworld Realtime TTS-2 Flash | 1,218 ±13 | 1,244 ±25 | 1,501 | 4.46 | 1 = | Headroom |
 Cartesia Sonic 3.5 | 1,190 ±11 | 1,203 ±24 | 1,574 | 4.41 | 15 | Visible delay |
 Fish Audio S2.1 Pro | 1,184 ±13 | 1,213 ±27 | n/a | 4.40 | 19 | Visible delay |
 Soniox TTS Real-Time v2 | 1,180 ±11 | 1,213 ±21 | 1,606 | 4.37 | 10 = | Visible delay |
 Gradium TTS | 1,174 ±13 | 1,241 ±26 | 1,592 | 4.51 | 1 = | Headroom |
 Smallest.ai Lightning V3.1 Pro (Jul 2026) | 1,169 ±12 | 1,189 ±24 | 1,447 | n/a | 20 | Visible delay |
 SpaceXAI Grok TTS | 1,155 ±13 | 1,193 ±27 | 1,495 | 4.17 | 18 | Visible delay |
 Murf AI Falcon 2 | 1,154 ±12 | 1,179 ±25 | n/a | n/a | 12 = | Visible delay |
 SpeechifyAI Simba 3.0 | 1,116 ±11 | 1,130 ±25 | n/a | n/a | 5 | Fast enough |
 ElevenLabs Flash v2.5 | 1,076 ±11 | 1,089 ±26 | 1,494 | 4.33 | 9 | Fast enough |
 Google Chirp 3: HD | 1,061 ±12 | 1,108 ±27 | n/a | n/a | 22 | Visible delay |
 Rime Coda | 1,054 ±12 | 1,108 ±26 | 1,535 | 4.42 | 16 = | Visible delay |
 Alibaba Qwen3 TTS Fast | n/a | n/a | 1,536 | n/a | 1 = | Headroom |
 Alibaba Qwen3 TTS 1.7b | n/a | n/a | n/a | n/a | 4 | Fast enough |
 Palabra TTS v1 | n/a | n/a | 1,520 | n/a | 7 | Fast enough |
 Soniox TTS RT v1 | n/a | n/a | 1,569 | n/a | 10 = | Visible delay |
 Rime Mist v3 | n/a | n/a | 1,428 | 4.34 | 12 = | Visible delay |
 Deepdub PhantomZ 4.0 | n/a | n/a | n/a | n/a | 16 = | Visible delay |
 Alibaba Qwen3 TTS Flash Realtime | n/a | n/a | n/a | n/a | 21 | Visible delay |
 Fish Audio S2.1 Pro Free | n/a | n/a | n/a | n/a | 23 | Visible delay |
Frequently asked questions
- Which multilingual TTS API sounds the most natural?
- On Artificial Analysis listener Elo, ElevenLabs Eleven v4 Turbo is highest among ranked models (1,333). Overlapping 95% intervals mean nearby models are tied. Speko and Cekura can order the same models differently.
- What is listener Elo?
- Artificial Analysis lets people compare two voices blind and turns the votes into an Elo score, with a 95% interval. A higher score means listeners preferred that model more often.
- Why not combine the naturalness scores?
- The sources test different models, texts and judges. Adding their numbers would imply a precision none of them has, so each stays in its own column.
- Does naturalness hold across languages?
- No source publishes naturalness per language. The Elo here is an overall figure, not a per-language one.