Data read October 27, 2026 · four public benchmarks
Which real-time TTS APIs stay fast beyond English?
Text-to-speech APIs for multilingual voice agents, ranked by speed first, with voice quality, accuracy and price side by side. Built from four public benchmarks, never merged into one score.
- 96
- models listed
- 23
- ranked on speed
- 15
- languages tested
- 35
- providers
Top 10 of the multilingual speed ranking
Ordered by speed tier, then P95 latency. Two P95 values within 10 ms share a rank, and tied models are listed by median latency. 3 models are in the Headroom tier, 6 in Fast enough and 14 in Visible delay.
| Rank | Model | Speed tier From the Coval P95 | Median (ms) Typical wait before the first sound | P95 (ms) 1 turn in 20 is slower | Listener Elo Artificial Analysis, with 95% interval | Price per 1M characters List price; a range when sources differ |
|---|---|---|---|---|---|---|
| 1 = |  Gradium TTS | Headroom | 53.7 | 104.3 | 1,174 ±13 | $36 to $57.80 |
| 1 = |  Inworld Realtime TTS-2 Flash | Headroom | 58.1 | 100.9 | 1,218 ±13 | $10.40 to $15 |
| 1 = |  Alibaba Qwen3 TTS Fast | Headroom | 62.9 | 98.8 | n/a | $10 |
| 4 |  Alibaba Qwen3 TTS 1.7b | Fast enough | 103.6 | 136.6 | n/a | n/a |
| 5 |  SpeechifyAI Simba 3.0 | Fast enough | 100.5 | 149.4 | 1,116 ±11 | $6.60 to $10 |
| 6 |  Inworld Realtime TTS-2 | Fast enough | 149.1 | 194.7 | 1,267 ±14 | $15 to $25 |
| 7 |  Palabra TTS v1 | Fast enough | 105.1 | 214.3 | n/a | $12 |
| 8 |  ElevenLabs Eleven v4 Turbo | Fast enough | 188.9 | 238.6 | 1,333 ±16 | $40 |
| 9 |  ElevenLabs Flash v2.5 | Fast enough | 191.0 | 255.1 | 1,076 ±11 | $40 to $50 |
| 10 = |  Soniox TTS RT v1 | Visible delay | 258.3 | 319.0 | n/a | $13 to $14.10 |
Pick the question that matters to your agent
- Ranking 23 multilingual real-time TTS APIs (text-to-speech) ranked by speed in the US, with latency, naturalness, accuracy and price side by side. Data read October 27, 2026.
- Latency Median and P95 time to first audio for 23 multilingual TTS APIs (text-to-speech), grouped into Headroom, Fast enough and Visible delay tiers. Coval 30-day data, US East.
- Multilingual Which multilingual TTS APIs (text-to-speech) are tagged multilingual by Coval, and in which of 15 languages Artificial Analysis or Speko tested each. Real-time speed shown beside.
- Voice agents Which real-time TTS APIs (text-to-speech) suit voice agents: P95 latency, streaming status, streamed-text input and customer-service listener Elo for 23 multilingual models.
- Pricing List price per 1M characters for 20 multilingual real-time TTS APIs (text-to-speech), as printed by four public benchmarks, with a range where they differ.
- Natural voices Which multilingual TTS APIs (text-to-speech) sound most natural: Artificial Analysis listener Elo, Speko blind votes and Cekura estimates for 23 real-time models, never added up.
- Accuracy How accurately multilingual TTS APIs (text-to-speech) read numbers, codes and hard passages: Cekura, Artificial Analysis and Speko results for 23 real-time models, side by side.
- Streaming Streaming TTS API (text-to-speech) comparison: streaming status, streamed-text input and Cekura first-audio times with and without streaming for 23 real-time multilingual models.
- Open weights 12 open-weights TTS models (text-to-speech) with their real-time speed tier, ranked first when multilingual and measured by Coval, plus languages tested.
- ElevenLabs alternatives 21 multilingual real-time TTS APIs (text-to-speech) to compare with ElevenLabs: speed tier, P95 latency, listener Elo and price, ranked by speed.
- Cartesia alternatives 21 multilingual real-time TTS APIs (text-to-speech) to compare with Cartesia: speed tier, P95 latency, listener Elo and price, ranked by speed.
- Inworld alternatives 21 multilingual real-time TTS APIs (text-to-speech) to compare with Inworld: speed tier, P95 latency, listener Elo and price, ranked by speed.
- Languages Which models were tested in each of 15 languages, from German to Telugu.
- Head-to-head One comparison page per pair, on latency, voice preference, accuracy and price.
How to read the ranking
- Speed sets the order, because a voice agent waits on the first sound before every answer.
- Naturalness, accuracy and price sit in their own columns. Nothing is added into one score.
- A figure comes from a named benchmark and a date. Where none holds it, the cell says n/a.
Frequently asked questions
- Which multilingual TTS API is the fastest in 2026?
- On the reading of October 27, 2026, Gradium TTS, Inworld Realtime TTS-2 Flash, and Alibaba Qwen3 TTS Fast share first place in the multilingual speed ranking, in the Headroom tier. The order among them can change from one reading to the next.
- What does this site do?
- It compares 96 text-to-speech APIs for real-time voice agents. It runs no tests: every figure is copied from Coval, Speko, Artificial Analysis or Cekura, with the source and the date it was read.
- Why is there no overall score?
- The four benchmarks test different models under different conditions. Speed sets the order, and voice quality, accuracy and price are shown beside it, never added together.
- Which models are ranked?
- 23 models: those Coval tags multilingual and measures. The other 73 are in the database with every measure we hold, but have no rank.
- Who publishes this ranking?
- Archipel Marketing. See the methodology page for how the data is read and how to ask for a correction.