Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Multilingual TTS APIs in 2026: 23 models and the languages each was tested in
23 models carry Coval's multilingual tag and a measured latency; 14 of them were also tested by Artificial Analysis or Speko in at least one language beyond English.
That gap is the main limit of this data: a multilingual tag says a vendor supports several languages, not that its speed or voice quality holds in each. Pick a language below to see who was tested in it.
How the rows are ordered. Same order as the ranking (speed tier, then P95, ties by median). Models are ordered by speed tier, then by P95 (the slow turn).
| Rank | Model | Languages tested | Languages tested | Coval tags | Speed tier From the Coval P95 | P95 (ms) 1 turn in 20 is slower |
|---|---|---|---|---|---|---|
| 1 = |  Gradium TTS | 4 | French, German, Portuguese, Spanish | multilingual, voice-cloning | Headroom | 104.3 |
| 1 = |  Inworld Realtime TTS-2 Flash | 9 | Arabic, Chinese, French, German, Hindi, Japanese, Portuguese, Spanish, Vietnamese | emotion-control, multilingual, voice-cloning | Headroom | 100.9 |
| 1 = |  Alibaba Qwen3 TTS Fast | 0 | none beyond English | multilingual | Headroom | 98.8 |
| 4 |  Alibaba Qwen3 TTS 1.7b | 0 | none beyond English | multilingual, voice-cloning | Fast enough | 136.6 |
| 5 |  SpeechifyAI Simba 3.0 | 4 | French, German, Portuguese, Spanish | emotion-control, multilingual, voice-cloning | Fast enough | 149.4 |
| 6 |  Inworld Realtime TTS-2 | 12 | Arabic, Chinese, Filipino, French, German, Hindi, Japanese, Korean, Portuguese, Spanish, Thai, Vietnamese | emotion-control, multilingual, voice-cloning | Fast enough | 194.7 |
| 7 |  Palabra TTS v1 | 0 | none beyond English | multilingual, voice-cloning | Fast enough | 214.3 |
| 8 |  ElevenLabs Eleven v4 Turbo | 9 | Arabic, Chinese, French, German, Hindi, Japanese, Portuguese, Spanish, Vietnamese | emotion-control, multilingual, voice-cloning | Fast enough | 238.6 |
| 9 |  ElevenLabs Flash v2.5 | 9 | Arabic, Chinese, French, German, Hindi, Japanese, Portuguese, Spanish, Vietnamese | multilingual, voice-cloning | Fast enough | 255.1 |
| 10 = |  Soniox TTS RT v1 | 3 | French, German, Spanish | multilingual, voice-cloning | Visible delay | 319.0 |
| 10 = |  Soniox TTS Real-Time v2 | 10 | Arabic, Chinese, Filipino, French, German, Hindi, Japanese, Norwegian, Spanish, Thai | multilingual, voice-cloning | Visible delay | 324.7 |
| 12 = |  Murf AI Falcon 2 | 0 | none beyond English | multilingual | Visible delay | 363.5 |
| 12 = |  Rime Mist v3 | 0 | none beyond English | multilingual | Visible delay | 359.9 |
| 14 |  Cartesia Sonic 3.6 | 9 | Arabic, Chinese, French, German, Hindi, Japanese, Portuguese, Spanish, Vietnamese | emotion-control, multilingual, voice-cloning | Visible delay | 374.4 |
| 15 |  Cartesia Sonic 3.5 | 15 | Arabic, Chinese, Filipino, French, German, Hindi, Japanese, Korean, Norwegian, Portuguese, Spanish, Tamil, Telugu, Thai, Vietnamese | emotion-control, multilingual, voice-cloning | Visible delay | 390.5 |
| 16 = |  Deepdub PhantomZ 4.0 | 0 | none beyond English | emotion-control, multilingual, voice-cloning | Visible delay | 430.0 |
| 16 = |  Rime Coda | 1 | Korean | multilingual | Visible delay | 428.1 |
| 18 |  SpaceXAI Grok TTS | 11 | Arabic, Chinese, Filipino, French, German, Hindi, Japanese, Portuguese, Spanish, Thai, Vietnamese | emotion-control, multilingual, voice-cloning | Visible delay | 473.9 |
| 19 |  Fish Audio S2.1 Pro | 2 | Chinese, German | emotion-control, multilingual, voice-cloning | Visible delay | 533.0 |
| 20 |  Smallest.ai Lightning V3.1 Pro (Jul 2026) | 0 | none beyond English | multilingual, voice-cloning | Visible delay | 703.3 |
| 21 |  Alibaba Qwen3 TTS Flash Realtime | 0 | none beyond English | multilingual | Visible delay | 785.8 |
| 22 |  Google Chirp 3: HD | 5 | Hindi, Japanese, Tamil, Telugu, Thai | multilingual | Visible delay | 917.2 |
| 23 |  Fish Audio S2.1 Pro Free | 0 | none beyond English | emotion-control, multilingual, voice-cloning | Visible delay | 1,137.9 |
Speed was measured in English only. A model with no language listed was tagged multilingual by Coval but not tested in another language by either source.
Frequently asked questions
- What makes a TTS API multilingual?
- Here, a model counts as multilingual when Coval tags it that way. This site does not test languages itself. For evidence in a given language, the languages tested column shows where Artificial Analysis or Speko ran the model in that language.
- Which multilingual TTS APIs were tested in the most languages?
- Starting with the most tested, Cartesia Sonic 3.5, Inworld Realtime TTS-2, and SpaceXAI Grok TTS lead the ranked models, with 15 languages for the first.
- Is multilingual TTS slower than English-only TTS?
- This data cannot say. Every latency figure is measured in English, so it does not show a per-language penalty. It only shows that 3 multilingual models stay under 120 ms P95 on the English test.
- Which languages can I compare?
- 15 languages have a page: each lists the models tested in it, ranked models first.