Text-to-Speech Model Leaderboard
Ranked by Elo from blind human listening votes. The default order is the Artificial Analysis (AA) TTS leaderboard, with TTS Arena V2 shown alongside; the two Elo scales use different baselines and cannot be compared directly.
Eleven v4 TurboElevenLabs · Elo 1336
#1 on TTS ArenaIInworld TTS MAXInworld · Elo 1548
Top from China
Qwen-Audio-3.1-TTS-PlusAlibaba · Elo 1307
Top open-weight
Fish Audio Drama 3 PreviewFish Audio · Elo 1248
Recently released
By release month listed on AA
Full leaderboard
24 modelsSorted by AA Elo (high to low) by default. Sort by TTS Arena Elo →
| # | Model | Artificial Analysis | TTS Arena V2 | Price $/1M chars | Released | Voices | ||
|---|---|---|---|---|---|---|---|---|
| Elo↑ | Samples | Elo | Votes | |||||
| 1 | Qwen3 TTSAlibabaCN | 930±13 | 3,113 | – | – | $13.8 | Jan 2026 | 8 |
| 2 | Qwen3 TTS FlashAlibabaCN | 942±14 | 2,220 | – | – | $11 | Sep 2025 | 4 |
| 3 |  T2A-01-TurboMiniMaxCN | 1031±11 | 5,000 | – | – | $30 | Jan 2025 | 8 |
| 4 | OpenAudio S1 MiniFish AudioOpenCN | 1045±20 | 1,682 | – | – | $15 | Jun 2025 | 8 |
| 5 |  MiMo-V2.5-TTSXiaomiCN | 1055±12 | 2,141 | – | – | – | – | 4 |
| 6 | T2A-01-HDMiniMaxCN | 1069±11 | 5,088 | – | – | $50 | Jan 2025 | 8 |
| 7 | OpenAudio S1Fish AudioCN | 1087±11 | 4,658 | 1491±28 | 608 | $15 | Jun 2025 | 8 |
| 8 | Speech-02-TurboMiniMaxCN | 1092±11 | 4,002 | 1495±21 | 1,231 | $60 | Mar 2025 | 4 |
| 9 | HSpeech 2.6HithinkCN | 1097±12 | 2,493 | 1491±30 | 587 | – | Jul 2026 | 8 |
| 10 |  Step Audio EditX (Mar 2026)StepFunOpenCN | 1100±12 | 2,359 | – | – | – | Feb 2026 | 2 |
| 11 | Fish Audio S2 ProFish AudioOpenCN | 1110±12 | 2,750 | – | – | $15 | Mar 2026 | 4 |
| 12 | Speech-02-HDMiniMaxCN | 1111±11 | 4,668 | 1504±21 | 1,243 | $100 | Mar 2025 | 4 |
| 13 | Speech 2.6 TurboMiniMaxCN | 1122±11 | 4,590 | 1514±31 | 542 | $60 | Oct 2025 | 8 |
| 14 | Speech 2.6 HDMiniMaxCN | 1130±11 | 4,316 | 1498±30 | 560 | $100 | Oct 2025 | 8 |
| 15 | Step TTS 2 (Mar 2026)StepFunCN | 1140±12 | 2,835 | – | – | $40 | Feb 2026 | 2 |
| 16 | Speech 2.8 TurboMiniMaxCN | 1156±10 | 5,016 | 1517±31 | 533 | $60 | Feb 2026 | 8 |
| 17 | Speech 2.8 HDMiniMaxCN | 1176±10 | 5,246 | 1532±29 | 503 | $100 | Feb 2026 | 8 |
| 18 | Fish Audio S2.1 Pro (Oct 2026)Fish AudioCN | 1186±14 | 2,123 | – | – | $15 | Oct 2026 | 8 |
| 19 |  Ming-3.0-ttsAnt GroupOpenCN | 1191±14 | 2,169 | – | – | – | Oct 2026 | 3 |
| 20 | StepAudio 2.5 TTS (Aug 2026)StepFunCN | 1214±14 | 2,254 | – | – | $85 | Aug 2026 | 8 |
| 21 | Fish Audio Drama 3 PreviewFish AudioOpenCN | 1248±15 | 2,237 | – | – | $15 | Oct 2026 | 8 |
| 22 | Qwen-Audio-3.0-TTS-PlusAlibabaCN | 1266±14 | 2,507 | – | – | $19.3 | Sep 2026 | 8 |
| 23 | Qwen-Audio-3.1-TTS-PlusAlibabaCN | 1307±16 | 2,030 | – | – | $19.3 | Sep 2026 | 8 |
| 24 | OpenAudio S2Fish AudioCN | – | – | 1515±28 | 603 | – | – | – |
Elo data: Artificial Analysis (artificialanalysis.ai) and TTS Arena V2 (TTS-AGI). Please credit the sources when citing.
Tracked, not yet rated on either leaderboard 4
Frequently asked-about open Chinese TTS models. Neither AA nor TTS Arena has blind-test scores for them yet, so they are not ranked; details come from their Hugging Face pages.
| Model | AA Elo | TTS Arena Elo | License | Released |
|---|---|---|---|---|
| Fun-CosyVoice3-0.5BAlibaba (FunAudioLLM)Open | – | – | Apache-2.0 | Dec 2025 |
| CosyVoice2-0.5BAlibaba (FunAudioLLM)Open | – | – | Apache-2.0 | Jul 2025 |
| Qwen3-TTS-12Hz-1.7BAlibabaOpen | – | – | Apache-2.0 | Jan 2026 |
| IndexTTS-2Bilibili (IndexTeam)Open | – | – | – | Jun 2025 |
Sources & methodology
- AAArtificial Analysis Text-to-Speech leaderboard
fetched Oct 25, 2026 · 98 models
- TATTS Arena V2 (TTS-AGI, Hugging Face)
fetched Oct 25, 2026 · 43 models
- Artificial Analysis TTS Arena: listeners compare two clips without knowing the models and vote; scores are Elo with Zonos-v0.1 fixed at 1000 as the anchor. ± is the half-width of the 95% confidence interval and “Samples” is the number of counted votes. Price is the official API list price converted to USD per 1M characters; release is by month.
- TTS Arena V2 (a TTS-AGI community project on Hugging Face): also blind pairwise-vote Elo; ± is its published uncertainty. Low-vote scores are marked “Prelim.” and suspended models are hidden. Its Elo baseline differs from AA (mostly 1350–1750), so compare each column only within itself.
- When both sources list the same model (e.g. MiniMax Speech 2.8 HD) they are merged into one row using a hand-maintained map that only joins exact name/version matches; anything ambiguous stays separate. “CN” is by vendor.
- New models stay visible: in the default order, up to 6 models that have no AA score but are active in the TTS Arena pool (anonymous codenames excluded) form a “No AA score yet” group inside the top 30 rows, ordered by TTS Arena Elo and marked “New” without a rank; they move into place once AA rates them.
- Missing values show “–”; we never estimate or invent scores or prices. Price and release month come only from AA’s paid data API; when the free API omits them, the last known value is kept and the hover text shows its as-of date.
- Synced automatically every day at 06:40 UTC. /api/public/tts-leaderboard.php
FAQ
What is the best text-to-speech model right now?
On AA’s blind-test Elo as of Oct 25, 2026, #1 is Eleven v4 Turbo by ElevenLabs (1336, ±17). On TTS Arena V2, #1 is Inworld TTS MAX (1548). Confidence intervals near the top often overlap, so gaps of a few dozen points are effectively ties.
Which Chinese TTS model ranks highest?
The highest Chinese model on AA is Qwen-Audio-3.1-TTS-Plus by Alibaba (1307, #3 overall); others on the board include Qwen-Audio-3.0-TTS-Plus, Fish Audio Drama 3 Preview, StepAudio 2.5 TTS (Aug 2026), Ming-3.0-tts, Fish Audio S2.1 Pro (Oct 2026). Open models such as CosyVoice have no scores on either leaderboard yet and are listed at the bottom.
Which open-weight TTS models can I run locally?
AA lists 18 open-weight models; the highest is Fish Audio Drama 3 Preview (1248). Use the “Open weights” filter to see them all. Licenses differ, so check each model page before commercial use.
Why are AA and TTS Arena scores so different?
They use different Elo baselines (AA anchors Zonos-v0.1 at 1000; TTS Arena is mostly 1350–1750) as well as different model pools, prompts and voters. Compare order within a column, never across columns.
Why do some new models sit near row 30 without a rank?
They already have TTS Arena scores but AA hasn’t rated them yet. So they aren’t buried at the bottom, the default order shows them as a group within the top 30 rows (marked “New”) until AA scores them.