llm leaderboard

Best AI for text to speech

Blind pairwise listening votes. Models still marked preliminary upstream are left out until their rating settles.

Aldena runs these models inside your team's projects. See what each one costs.

23 of 23 ranked models
Ranked models
rankmodelvendorcompositebenchmarkselo
1
Luck Dolphinluck-dolphin/luck-dolphin
Luck Dolphin66.71 of 1 benchmarks1684
2
Inworld TTS (max)inworld/inworld:max
Inworld65.11 of 1 benchmarks1567
3
Inworld TTS 1.5 MAXinworld/inworld-max-1.5
Inworld63.61 of 1 benchmarks1565
4
CastleFlow v1.0async/async-1
Async62.11 of 1 benchmarks1557
5
Hume Octavehume/hume-octave
Hume59.91 of 1 benchmarks1551
6
Papla P1papla/papla-p1
Papla59.91 of 1 benchmarks1551
7
Inworld TTSinworld/inworld
Inworld57.61 of 1 benchmarks1543
8
OpenAudio S2lanternfish/lanternfish-2
Lanternfish56.11 of 1 benchmarks1542
9
Eleven Multilingual v2elevenlabs/eleven-multilingual-v2
ElevenLabs54.51 of 1 benchmarks1507
10
Eleven Turbo v2.5elevenlabs/eleven-turbo-v2.5
ElevenLabs53.01 of 1 benchmarks1504
11
Eleven Flash v2.5elevenlabs/eleven-flash-v2.5
ElevenLabs50.81 of 1 benchmarks1503
12
MiniMax Speech 02 HDminimax/minimax-speech-02-hd
MiniMax50.81 of 1 benchmarks1503
13
MiniMax Speech 02 Turbominimax/minimax-speech-02-turbo
MiniMax48.51 of 1 benchmarks1482
14
Typecast SSFM 3.0typecast/typecast
Typecast47.01 of 1 benchmarks1475
15
Kokoro v1.0hexgrad/kokoro-v1
Hexgrad45.51 of 1 benchmarks1469
16
OpenAudio S1lanternfish/lanternfish-1
Lanternfish43.91 of 1 benchmarks1465
17
Chatterboxresemble-ai/chatterbox
Resemble AI42.41 of 1 benchmarks1460
18
Magpie Research Previewmagpie/magpie-rp
Magpie40.91 of 1 benchmarks1456
19
Hithink Speech 2.6hithink/hithink-speech-2.6
Hithink39.41 of 1 benchmarks1454
20
NeuTTS Maxneuphonic/neuphonic
Neuphonic37.91 of 1 benchmarks1389
21
Maya 1maya-research/maya1
Maya Research36.41 of 1 benchmarks1383
22
Magpie Multilingualmagpie/magpie
Magpie34.81 of 1 benchmarks1372
23
Wordcab TTSwordcab/wordcab
Wordcab33.31 of 1 benchmarks1331
How this ranks

Every benchmark value becomes a percentile among the models that have it, so accuracy scores, Elo ratings and word error rates compare without hand-tuned scaling. Metrics where lower is better are inverted first. Raw values are never summed or averaged across benchmarks. A model's mean percentile is then shrunk toward the mean of the models that were broadly benchmarked, so a model tested twice cannot outrank a broadly tested one on two lucky results. Turning a data source off runs that same ranking code again in your browser over the sources you left on.

A model scored on fewer than 1 of the 1 ranked benchmarks in this category still ranks here, on the benchmarks it does have, and its row carries a partial coverage mark. On an equal score it sits under the model that earned the same number across more of the board.

Data sources

Turn a source off to drop every benchmark it feeds and re-rank the board in your browser. Your choice follows you across the leaderboard pages.