Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Real-time TTS API latency in 2026: 23 multilingual models by speed tier
Latency here is the time between sending text and hearing the first sound, measured by Coval over 30 days from the US East. 3 models sit in the Headroom tier (P95 under 120 ms), 6 in Fast enough (120 to 293 ms) and 14 in Visible delay (above 293 ms).
Low latency matters most to voice agents, where the voice becomes part of the pause before every answer. The slowest ranked model, Fish Audio S2.1 Pro Free, has a P95 of 1,137.9 ms.
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.
Headroom (3 models)
P95 under 120 ms: room left in the turn for everything else.
| Rank | Model | Median (ms) Typical wait before the first sound | P90 (ms) | P95 (ms) 1 turn in 20 is slower | Std dev (ms) | Samples Calls in the 30-day window |
|---|---|---|---|---|---|---|
| 1 = |  Gradium TTS | 53.7 | 92.3 | 104.3 | 21.9 | 2,808 |
| 1 = |  Inworld Realtime TTS-2 Flash | 58.1 | 83.7 | 100.9 | 874.1 | 2,612 |
| 1 = |  Alibaba Qwen3 TTS Fast | 62.9 | 80.7 | 98.8 | 20.6 | 2,808 |
Fast enough (6 models)
P95 from 120 to 293 ms: enough for most agents.
| Rank | Model | Median (ms) Typical wait before the first sound | P90 (ms) | P95 (ms) 1 turn in 20 is slower | Std dev (ms) | Samples Calls in the 30-day window |
|---|---|---|---|---|---|---|
| 4 |  Alibaba Qwen3 TTS 1.7b | 103.6 | 129.3 | 136.6 | 17.8 | 2,586 |
| 5 |  SpeechifyAI Simba 3.0 | 100.5 | 134.0 | 149.4 | 83.1 | 2,805 |
| 6 |  Inworld Realtime TTS-2 | 149.1 | 181.9 | 194.7 | 504.6 | 2,612 |
| 7 |  Palabra TTS v1 | 105.1 | 153.4 | 214.3 | 67.1 | 2,318 |
| 8 |  ElevenLabs Eleven v4 Turbo | 188.9 | 221.9 | 238.6 | 46.7 | 2,466 |
| 9 |  ElevenLabs Flash v2.5 | 191.0 | 232.7 | 255.1 | 33.9 | 2,798 |
Visible delay (14 models)
P95 above 293 ms: the voice is a visible part of the delay.
| Rank | Model | Median (ms) Typical wait before the first sound | P90 (ms) | P95 (ms) 1 turn in 20 is slower | Std dev (ms) | Samples Calls in the 30-day window |
|---|---|---|---|---|---|---|
| 10 = |  Soniox TTS RT v1 | 258.3 | 307.0 | 319.0 | 42.9 | 2,372 |
| 10 = |  Soniox TTS Real-Time v2 | 264.4 | 311.3 | 324.7 | 47.9 | 2,365 |
| 12 = |  Murf AI Falcon 2 | 171.3 | 352.7 | 363.5 | 105.8 | 2,805 |
| 12 = |  Rime Mist v3 | 275.2 | 348.0 | 359.9 | 45.2 | 254 |
| 14 |  Cartesia Sonic 3.6 | 285.8 | 350.6 | 374.4 | 70.5 | 2,807 |
| 15 |  Cartesia Sonic 3.5 | 279.3 | 349.6 | 390.5 | 122.5 | 2,807 |
| 16 = |  Deepdub PhantomZ 4.0 | 183.4 | 349.7 | 430.0 | 327.9 | 2,796 |
| 16 = |  Rime Coda | 317.5 | 390.5 | 428.1 | 54.7 | 254 |
| 18 |  SpaceXAI Grok TTS | 329.8 | 427.3 | 473.9 | 155.4 | 2,803 |
| 19 |  Fish Audio S2.1 Pro | 328.1 | 445.6 | 533.0 | 108.3 | 2,806 |
| 20 |  Smallest.ai Lightning V3.1 Pro (Jul 2026) | 360.5 | 586.0 | 703.3 | 139.9 | 2,797 |
| 21 |  Alibaba Qwen3 TTS Flash Realtime | 726.8 | 768.7 | 785.8 | 44.1 | 2,807 |
| 22 |  Google Chirp 3: HD | 534.1 | 847.1 | 917.2 | 220.0 | 2,808 |
| 23 |  Fish Audio S2.1 Pro Free | 793.8 | 1,011.4 | 1,137.9 | 327.6 | 2,792 |
Fluxions vui has the lowest P95 Coval measures overall (75.0 ms), but Coval does not tag it multilingual and no source tested it outside English, so it is not in this multilingual ranking.
Frequently asked questions
- What latency counts as real-time for a TTS API?
- Voice agents need the first sound quickly, and the slow turns matter as much as the typical one. This site uses P95 under 120 ms as Headroom, 120 to 293 ms as Fast enough, and above 293 ms as Visible delay. The breaks are placed in the widest gaps between consecutive P95 values.
- What is the difference between median and P95 latency?
- The median (P50) is the typical wait: half of the calls are faster. P95 is the slow turn: one call in twenty is slower. A model can lead on one and trail on the other, so both are shown.
- Which multilingual TTS API has the lowest latency?
- In ranking order, the first four are Gradium TTS, Inworld Realtime TTS-2 Flash, Alibaba Qwen3 TTS Fast, and Alibaba Qwen3 TTS 1.7b; the lowest P95 among ranked models is 98.8 ms (Alibaba Qwen3 TTS Fast). The 3 models at rank 1 are within 10 ms of each other on P95, so their order can move from one reading to the next.
- Where is TTS latency measured?
- Coval calls each hosted API every 30 minutes from the US East and publishes percentiles over 30 days. Latency from Europe or Asia is not covered by any public source we use.