Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Best multilingual TTS API in 2026: 23 real-time models ranked by speed
On the reading of October 27, 2026, Gradium TTS, Inworld Realtime TTS-2 Flash, and Alibaba Qwen3 TTS Fast share first place among the 23 multilingual real-time TTS APIs we rank, all in the Headroom speed tier with a P95 under 120 ms.
The ranking is a speed order, not a quality score: it never adds speed to voice quality or price. Quality, accuracy and price sit in the columns beside it, and the pages linked below rank on each of them. A model is ranked when Coval, a public benchmark, both tags it multilingual and measures its latency.
How the rows are ordered. Models are ordered by speed tier, then by P95 (the slow turn). Two P95 values within 10 ms share a rank, and tied models are listed by median latency, a rule applied to every tie. The "Best for" column is computed from these measures, not declared by the vendor.
| Rank | Model | Speed tier From the Coval P95 | Median (ms) Typical wait before the first sound | P95 (ms) 1 turn in 20 is slower | Languages tested | Listener Elo Artificial Analysis, with 95% interval | Price per 1M characters List price; a range when sources differ | Best for Computed: where the model leads among ranked models |
|---|---|---|---|---|---|---|---|---|
| 1 = |  Gradium TTS | Headroom | 53.7 | 104.3 | 4 | 1,174 ±13 | $36 to $57.80 | Lowest median latency |
| 1 = |  Inworld Realtime TTS-2 Flash | Headroom | 58.1 | 100.9 | 9 | 1,218 ±13 | $10.40 to $15 | Rated for knowledge sharing and assistants |
| 1 = |  Alibaba Qwen3 TTS Fast | Headroom | 62.9 | 98.8 | 0 | n/a | $10 | Lowest P95 latency |
| 4 |  Alibaba Qwen3 TTS 1.7b | Fast enough | 103.6 | 136.6 | 0 | n/a | n/a | Rated for voice agents |
| 5 |  SpeechifyAI Simba 3.0 | Fast enough | 100.5 | 149.4 | 4 | 1,116 ±11 | $6.60 to $10 | Lowest list price |
| 6 |  Inworld Realtime TTS-2 | Fast enough | 149.1 | 194.7 | 12 | 1,267 ±14 | $15 to $25 | Best reading accuracy (Cekura) |
| 7 |  Palabra TTS v1 | Fast enough | 105.1 | 214.3 | 0 | n/a | $12 | Rated for voice agents |
| 8 |  ElevenLabs Eleven v4 Turbo | Fast enough | 188.9 | 238.6 | 9 | 1,333 ±16 | $40 | Highest listener Elo; Highest naturalness (Speko) |
| 9 |  ElevenLabs Flash v2.5 | Fast enough | 191.0 | 255.1 | 9 | 1,076 ±11 | $40 to $50 | Rated for knowledge sharing and assistants |
| 10 = |  Soniox TTS RT v1 | Visible delay | 258.3 | 319.0 | 3 | n/a | $13 to $14.10 | Visible delay tier, no measure where it leads |
| 10 = |  Soniox TTS Real-Time v2 | Visible delay | 264.4 | 324.7 | 10 | 1,180 ±11 | $13 to $14.81 | Rated for knowledge sharing and assistants |
| 12 = |  Murf AI Falcon 2 | Visible delay | 171.3 | 363.5 | 0 | 1,154 ±12 | $10 | Rated for knowledge sharing and assistants |
| 12 = |  Rime Mist v3 | Visible delay | 275.2 | 359.9 | 0 | n/a | $30 | Visible delay tier, no measure where it leads |
| 14 |  Cartesia Sonic 3.6 | Visible delay | 285.8 | 374.4 | 9 | 1,292 ±14 | $39.20 to $50 | Rated for knowledge sharing and assistants |
| 15 |  Cartesia Sonic 3.5 | Visible delay | 279.3 | 390.5 | 15 | 1,190 ±11 | $39.20 to $50 | Most languages tested |
| 16 = |  Deepdub PhantomZ 4.0 | Visible delay | 183.4 | 430.0 | 0 | n/a | n/a | Visible delay tier, no measure where it leads |
| 16 = |  Rime Coda | Visible delay | 317.5 | 428.1 | 1 | 1,054 ±12 | $50 | Rated for knowledge sharing and assistants |
| 18 |  SpaceXAI Grok TTS | Visible delay | 329.8 | 473.9 | 11 | 1,155 ±13 | $15 | Rated for knowledge sharing and assistants |
| 19 |  Fish Audio S2.1 Pro | Visible delay | 328.1 | 533.0 | 2 | 1,184 ±13 | $15 | Rated for knowledge sharing and assistants |
| 20 |  Smallest.ai Lightning V3.1 Pro (Jul 2026) | Visible delay | 360.5 | 703.3 | 0 | 1,169 ±12 | $19.50 | Rated for knowledge sharing and assistants |
| 21 |  Alibaba Qwen3 TTS Flash Realtime | Visible delay | 726.8 | 785.8 | 0 | n/a | $13 | Visible delay tier, no measure where it leads |
| 22 |  Google Chirp 3: HD | Visible delay | 534.1 | 917.2 | 5 | 1,061 ±12 | $30 | Rated for knowledge sharing and assistants |
| 23 |  Fish Audio S2.1 Pro Free | Visible delay | 793.8 | 1,137.9 | 0 | n/a | n/a | Visible delay tier, no measure where it leads |
73 other models are in the database but not ranked here, because Coval does not tag them multilingual or does not measure them. They keep a page with every measure we hold. See the methodology for the inclusion rule.
Frequently asked questions
- What is the best multilingual TTS API in 2026?
- On speed, Gradium TTS, Inworld Realtime TTS-2 Flash, and Alibaba Qwen3 TTS Fast share first place on the October 27, 2026 reading, each with a P95 of 104.3 ms or less. On listener preference the order differs: among ranked models the highest Artificial Analysis Elo is ElevenLabs Eleven v4 Turbo (1,333). Which is best depends on whether your agent is limited by delay, voice quality, accuracy or price.
- How is this ranking built?
- Models are grouped into speed tiers by their Coval P95 latency, then ordered by P95 inside each tier. Two P95 values within 10 ms are a tie, and tied models are shown by median latency. No benchmark ranks or scores are added together.
- Which multilingual TTS APIs are the fastest?
- In ranking order (speed tier, then P95, tied models by median latency), the first five are Gradium TTS, Inworld Realtime TTS-2 Flash, Alibaba Qwen3 TTS Fast, Alibaba Qwen3 TTS 1.7b, and SpeechifyAI Simba 3.0. Speed is measured in English from the US East, so it says how fast a model starts, not how fast it is in a given language.
- Does the ranking measure quality in each language?
- No. No public benchmark publishes naturalness or accuracy per language. The languages tested column only says in which languages Artificial Analysis or Speko tested the model. See the language pages for who was tested in each of the 15 languages.
- How many TTS models does this site cover?
- 96 models from 35 providers are published, of which 23 are ranked on the multilingual speed order.