Speech synthesis ranked with the voice held constant
Compare controlled-voice TTS rankings for a consistent narrator and plan a blind listening check using a consented voice.

👁 111 reads
A museum audio guide needs the same narrator in the new wing as in the old one. For that job I would start with controlled voice. Holding the speaker reference steady makes it easier to hear what the system does to pauses, emphasis and vowel sounds.
What the board measures
HeyGen Voice leads this track while Eleven v4 Turbo heads the provider-voice track. These tables answer different casting questions. For the museum, check whether a listener still recognizes the approved speaker when a street address interrupts a sentence. Keep the pronunciation score separate from voice resemblance.
The November 17, 2026 snapshot
The table covers the top 25 and additional recently added entries captured in the source. Prices are the evaluator’s reported USD benchmark estimates. A blank price means the evaluator did not supply one.
| Rank and model | Elo and votes | Benchmark price |
|---|---|---|
| 1. HeyGen Voice | 1192 (2,807 votes) | $30.00 /1M chars |
| 2. Simba 3.5 | 1189 (1,990 votes) | $6.58 /1M chars |
| 3. HeyGen Voice Turbo | 1178 (2,596 votes) | $4.00 /1M chars |
| 4. Qwen-Audio-3.1-TTS-Plus | 1175 (3,122 votes) | $19.31 /1M chars |
| 5. Qwen-Audio-3.1-TTS-Flash | 1171 (1,763 votes) | $19.31 /1M chars |
| 6. Eleven v4 Turbo | 1156 (3,469 votes) | $40.00 /1M chars |
| 7. Eleven v4 | 1155 (3,352 votes) | $80.00 /1M chars |
| 8. Sonic 3.6 | 1150 (4,217 votes) | $49.00 /1M chars |
| 9. Realtime TTS-2 | 1145 (3,936 votes) | $20.83 /1M chars |
| 10. Qwen-Audio-3.0-TTS-Plus | 1133 (3,716 votes) | $19.31 /1M chars |
| 11. Sonic 3.5 | 1118 (6,212 votes) | $49.00 /1M chars |
| 12. Lightning V3.1 Pro (Oct 2026) | 1101 (2,377 votes) | $19.50 /1M chars |
| 13. AssemblyAI-TTS-1 | 1089 (2,267 votes) | $15.00 /1M chars |
| 14. Simba 3.2 | 1081 (3,761 votes) | $6.58 /1M chars |
| 15. StepAudio 2.5 TTS | 1076 (4,635 votes) | $85.00 /1M chars |
| 16. Eleven v3 | 1075 (5,253 votes) | $100.00 /1M chars |
| 17. Fish Audio Drama 3 | 1070 (2,374 votes) | $15.00 /1M chars |
| 18. Ming-3.0-tts | 1070 (2,675 votes) | N/A |
| 19. Voxtral 3 TTS | 1068 (2,570 votes) | $13.00 /1M chars |
| 20. Realtime TTS-2 Flash | 1066 (4,019 votes) | $10.42 /1M chars |
| 21. v3 Conversational | 1066 (3,510 votes) | $50.00 /1M chars |
| 22. SpaceXAI TTS | 1064 (4,210 votes) | $15.00 /1M chars |
| 23. Fish Audio Drama 3 Preview | 1059 (2,928 votes) | $15.00 /1M chars |
| 24. Gradium TTS (Sep 2026) | 1055 (3,059 votes) | $47.78 /1M chars |
| 25. Gemini 3.8 Flash-Lite TTS | 1055 (3,844 votes) | $11.03 /1M chars |
Reading the result
Preference and cost. The score comes from the listening or viewing comparison. The price column describes a separate estimate. For a working budget I would record actual spend across all attempts and divide by the number of accepted deliverables.
For the next step, see the Quantslant voice cloning guide.
Changes since the October archive
The October 13 crawl contains 43 entries on this track. The November 17 board contains 51. The following labels occur only in the later capture. This comparison does not establish the exact release day or rule out a renamed entry.
- AssemblyAI-TTS-1
- Fish Audio Drama 3
- HeyGen Voice Turbo
- Lightning V3.1 Pro (Oct 2026)
- Ming-3.0-tts
- Qwen-Audio-3.1-TTS-Flash
- Simba 3.5
- Voxtral 3 TTS
01Eleven v4 Review: Numbers, Dates & Emotion Tags in EnglishEleven v4 tops the Artificial Analysis TTS leaderboard with 1319 Elo. We tested it in English: 18/18 numbers and abbreviations correct, 0.25s to first audio in Turbo.Read →
02ElevenLabs Voice Cloning: clones from 10 s, 30 s and 2 minWe cloned a synthetic voice from 10 s, 30 s and 2 minutes of audio. The 10-second clone reached 0.992 similarity, the 2-minute clone 0.996.Read →Sources
- https://artificialanalysis.ai/text-to-speech/leaderboard/controlled-voice (accessed November 17, 2026)
- https://artificialanalysis.ai/text-to-speech/leaderboard/controlled-voice?ref=aiposthub.com (accessed November 17, 2026)
More on this from QUANTSLANT
benchmarks·Cognition models, access and benchmark guideCognition model and access guide with a practical US-English evaluation brief and links to publisher materials.