September 22, 2026
English does not explain TTS quality: Sonic leads in 8 of 9 languages
On September 22, Cartesia's Sonic family led 8 of 9 new language-specific TTS rankings. Sonic 3.6 took first place in 7 languages, while Inworld's Realtime TTS-2 led in Mandarin Chinese.
Artificial Analysis
@artificialanlys
Introducing the multilingual Text to Speech Arena rankings: we compared leading TTS models in 9 languages beyond English 🇯🇵 🇨🇳 🇮🇳 🇪🇸 🇩🇪 🇫🇷 🇵🇹 🇻🇳 🇸🇦 TTS quality varies significantly across languages: a model that works well in English will not necessarily handle pronunciation, pace, intonation, and naturalness as well in other languages. We expanded the Artificial Analysis Controlled Voice Arena to 9 new languages: Japanese 🇯🇵, Mandarin Chinese 🇨🇳, Hindi 🇮🇳, Spanish 🇪🇸, German 🇩🇪, French 🇫🇷, Portuguese 🇵🇹, Vietnamese 🇻🇳, and Arabic 🇸🇦. Each model is evaluated on standardized cloned voices. Prompts are written by native speakers of each language, and preference votes are collected from native speakers. Elo is calculated separately for each language, so every ranking reflects the preferences of its native speakers. Key results: ➤ Cartesia's Sonic family leads 8 of 9 new language-specific rankings. Sonic 3.6 took first place in 7 languages, while Sonic 3.5 leads in Portuguese. Inworld's Realtime TTS-2 took first place in Mandarin Chinese. ➤ ElevenLabs' Eleven v3 family made the top three in 7 of 9 new languages, thanks to Eleven v3 and Eleven v3 Conversational. ➤ Inworld's Realtime TTS-2 leads in Mandarin Chinese with 1185 Elo. Sonic 3.6 scored 1146, and StepFun's StepAudio 2.5 TTS scored 1130. Realtime TTS-2 also ranks first in US English. ➤ More than 100 thousand human votes were collected across the 9 new language-specific rankings. Each language ranks between 15 and 24 public models. See the language-by-language results below ⬇️
· 23.6K views
Artificial Analysis now compares TTS separately in Japanese, Mandarin Chinese, Hindi, and 6 other languages. A model may sound good in English but handle pronunciation, pace, and intonation differently in another language.
Choose a voice by language. More than 100 thousand votes from native speakers created separate rankings, with 15 to 24 public models competing in each language. A single overall leader is no longer enough for a multilingual app.
Each of the 9 languages received a separate ranking with independent Elo.
