Best text-to-speech models
Voice models ranked by head-to-head human preference.
ranked by elo · 96 models · updated Oct 13 · 51–96
| # | Model | Creator | Elo |
|---|---|---|---|
| 51 | T2A-01-HD | MiniMax | 1,068 |
| 52 | Maya 2 Flash | Maya Research | 1,067 |
| 53 | Polly Generative | Amazon | 1,066 |
| 54 | Kokoro 82M v1.0 | Kokoro | 1,063 |
| 55 | Magpie-Multilingual 357M (Feb 2026) | NVIDIA | 1,063 |
| 56 | Sonic 3 | Cartesia | 1,060 |
| 57 | Chirp 3: HD | 1,060 | |
| 58 | Gemini 2.5 Flash TTS (Dec 2025) | 1,060 | |
| 59 | Coda | Rime | 1,059 |
| 60 | MiMo-V2.5-TTS | Xiaomi | 1,056 |
| 61 | Flux TTS | Deepgram | 1,053 |
| 62 | Octave 2 | Hume AI | 1,052 |
| 63 | Async Flash v1.0 | async | 1,050 |
| 64 | Maya1 | Maya Research | 1,047 |
| 65 | Maya 2 Global | Maya Research | 1,046 |
| 66 | OpenAudio S1 Mini | Fish Audio | 1,045 |
| 67 | Octave TTS | Hume AI | 1,041 |
| 68 | Higgs Audio V3 TTS | Boson AI | 1,038 |
| 69 | Bland Speech v3 | Bland AI | 1,037 |
| 70 | Azure Neural | Microsoft | 1,036 |
| 71 | Gemini 2.5 Pro (Dec 2025) | 1,033 | |
| 72 | MAI-Voice-1 | Microsoft | 1,033 |
| 73 | Polly Long-Form | Amazon | 1,030 |
| 74 | T2A-01-Turbo | MiniMax | 1,029 |
| 75 | Chatterbox | Resemble AI | 1,027 |
| 76 | Lightning v3.1 | Smallest.ai | 1,025 |
| 77 | MiMo-V2-TTS | Xiaomi | 1,024 |
| 78 | Raon SpeechLM | Krafton | 1,004 |
| 79 | Zonos-v0.1 | Zyphra | 1,000 |
| 80 | Murf Speech Gen 2 | Murf AI | 986 |
| 81 | VibeVoice 1.5B | Microsoft | 957 |
| 82 | OpenVoice v2 | OpenVoice | 955 |
| 83 | Qwen3 TTS Flash | Alibaba | 943 |
| 84 | Neuphonic TTS | Neuphonic | 942 |
| 85 | Kugel 3 | KugelAudio | 935 |
| 86 | Qwen3 TTS | Alibaba | 932 |
| 87 | WaveNet | 918 | |
| 88 | XTTS v2 | Coqui | 916 |
| 89 | Mist V2 | Rime | 908 |
| 90 | Polly Neural | Amazon | 896 |
| 91 | Neural2 | 895 | |
| 92 | StyleTTS 2 | StyleTTS | 895 |
| 93 | Standard | 888 | |
| 94 | Noiz TTS | Noiz | 879 |
| 95 | MetaVoice v1 | MetaVoice | 848 |
| 96 | Polly Standard | Amazon | 832 |
what elo means
An Elo rating from head-to-head votes: two outputs are shown side by side and people pick the better one. Higher is better, and a gap of about 100 points means the higher model wins roughly two thirds of the time.