Text-to-Speech
Multilingual
Naturalness beyond English — Spanish, German, French, Arabic, Filipino and Norwegian, each from a blind A/B listening study on Prolific. Internal MOS (Bradley-Terry) with a 95% CI; overlapping intervals are a statistical tie.
| # | System | MOS · higher ▸ | 95% CI |
|---|---|---|---|
| 1 | eleven_v3ElevenLabs | 4.68 | 4.45–4.90 |
| 2 | sonic-3.5Cartesia | 4.12 | 3.77–4.44 |
| 3 | grok-ttsxAI Grok | 3.64 | 3.31–3.97 |
| 4 | inworld-tts-2Inworld | 3.48 | 3.21–3.77 |
| 5 | octave-2Hume | 3.33 | 2.93–3.70 |
| 6 | tts-rt-v1Soniox | 3.12 | 2.76–3.55 |
| 7 | lightning_v3.1Smallest | 2.93 | 2.61–3.26 |
| 8 | speech-2.8-hdMiniMax | 2.80 | 2.52–3.13 |
| 9 | qwen3-tts-flashQwen | 2.69 | 2.42–2.97 |
| 10 | gpt-4o-mini-ttsOpenAI | 2.55 | 2.31–2.75 |
| 11 | arcanav3Rime | 1.63 | 1.36–1.95 |
| 12 | aura-2Deepgram | 1.34 | 1.13–1.64 |
