Real-time TTS · speed only · 2026
Best real-time TTS API, ranked on measured speed
This ranking looks at one thing only: how fast each text-to-speech API starts speaking. It orders the 44 models whose latency is measured by Coval or Speko on P50 and P99 time-to-first-audio (TTFA) and latency consistency, nothing else. Voice quality, accuracy and price are not counted here; for the realtime choice that weighs all four, see the overall TTS benchmark.
Coval 30-day window ending October 1, 2026 · Speko US East P50 · updated after each benchmark capture
| # | Model | Speed /60 | P50 TTFA (Coval) | P99 TTFA (Coval) | Consistency (std dev) | P50 TTFA (Speko) |
|---|---|---|---|---|---|---|
| 1 | Gradium TTS Beta Gradium | 58.7 | 47.2 ms | 108.8 ms | 16.6 ms | 100 ms |
| 2 | Alibaba Qwen3 TTS Fast Alibaba (Qwen) | 57.7 | 64.6 ms | 125.3 ms | 14.6 ms | 64 ms |
| 3 | Palabra TTS v1 Palabra | 49.9 | 105 ms | 299.8 ms | 42.4 ms | 72 ms |
| 4 | ElevenLabs Flash v2.5 ElevenLabs | 45.7 | 184.8 ms | 332.4 ms | 33.2 ms | — |
| 5 | Fluxions vui Fluxions | 44.6 | 50.2 ms | 128.1 ms | 241.6 ms | — |
| 6 | Alibaba Qwen3 TTS 1.7b Alibaba (Qwen) | 43.9 | 103.5 ms | 165.4 ms | 37.5 ms | — |
| 7 | Inworld Realtime TTS-2 Flash Inworld | 43.5 | 69.4 ms | 175.2 ms | 373.3 ms | 82 ms |
| 8 | Soniox TTS Real-Time v2 Soniox | 40.1 | 245.6 ms | 343.4 ms | 49.8 ms | 381 ms |
| 9 | Rime Mist v3 Rime | 40 | 254.8 ms | 364.3 ms | 26.1 ms | — |
| 10 | Gradium TTS (Aug 2026) Gradium | 39.7 | 213.5 ms | 382.5 ms | 70.2 ms | 272 ms |
| 11 | Soniox TTS RT v1 Soniox | 37.9 | 247.2 ms | 339.8 ms | 104.8 ms | 362 ms |
| 12 | Inworld Realtime TTS-2 Inworld | 37.8 | 163.1 ms | 293.7 ms | 567.8 ms | 116 ms |
| 13 | Cartesia Sonic 3.5 Cartesia | 37.1 | 280 ms | 421.8 ms | 45.6 ms | 121 ms |
| 14 | Deepgram Flux TTS Deepgram | 36.4 | 193.9 ms | 550.3 ms | 131 ms | 106 ms |
| 15 | Deepdub Phantom Z 3.4 conversational Deepdub | 33.3 | 261.3 ms | 427.4 ms | 80.8 ms | — |
| 16 | Rime Coda Rime | 32.4 | 294.2 ms | 438.1 ms | 46 ms | — |
| 17 | ElevenLabs Eleven v3 Conversational ElevenLabs | 29.9 | 323.1 ms | 497.7 ms | 52.9 ms | 266 ms |
| 18 | Deepgram Aura 2 Deepgram | 27.9 | 287.9 ms | 576 ms | 140.8 ms | 125 ms |
| 19 | Cartesia Sonic 3.6 Cartesia | 21.5 | 347.9 ms | 840 ms | 117.6 ms | 120 ms |
| 20 | Smallest.ai Lightning V3.1 Pro (Jul 2026) Smallest.ai | 21.4 | 337.1 ms | 793.8 ms | 117.4 ms | — |
| 21 | SpeechifyAI Simba 3.2 Speechify | 19.6 | 362.7 ms | 921.5 ms | 388.2 ms | 85 ms |
| 22 | SpaceXAI Grok TTS xAI | 19.3 | 386.6 ms | 542.4 ms | 163.1 ms | 272 ms |
| 23 | Fish Audio S2.1 Pro Fish Audio | 18.6 | 300.1 ms | 885.9 ms | 193.9 ms | — |
| 24 | Fish Audio OpenAudio S1 Fish Audio | 17.1 | 364.1 ms | 789.6 ms | 163.5 ms | — |
| 25 | Murf AI Falcon 2 Murf AI | 14.8 | 557.5 ms | 860.9 ms | 141.3 ms | — |
| 26 | SpeechifyAI Simba 3.0 Speechify | 14.3 | 390.6 ms | 906.3 ms | 498.6 ms | — |
| 27 | Google Chirp 3: HD Google | 10 | 515.2 ms | 1282.2 ms | 225.8 ms | — |
| 28 | Deepgram Aura 2 En Deepgram | 6.9 | 520.2 ms | 1696.9 ms | 326.6 ms | — |
| 29 | Alibaba Qwen3 TTS Flash Realtime Alibaba (Qwen) | 4.3 | 708.8 ms | 4208.7 ms | 566.7 ms | — |
| 30 | Nari Qwen3-TTS Alibaba (Qwen) | 2.7 | — | — | — | 75 ms |
| 31 | Maya 2 Native Maya Research | 2.3 | — | — | — | 100 ms |
| 32 | Smallest.ai Lightning v3.1 Smallest.ai | 1.6 | — | — | — | 173 ms |
| 33 | ElevenLabs Eleven v4 Turbo ElevenLabs | 1.5 | — | — | — | 205 ms |
| 34 | Rime arcanav3 Rime | 1.4 | — | — | — | 238 ms |
| 35 | OpenAI GPT-4o mini TTS OpenAI | 1.3 | 567.4 ms | 5896.6 ms | 1575.6 ms | 691 ms |
| 36 | Fish Audio S2.1 Pro Free Fish Audio | 1 | 405.6 ms | 7919.1 ms | 2855.9 ms | — |
| 36 | MiniMax Speech 2.8 HD MiniMax | 1 | — | — | — | 294 ms |
| 38 | Bland Speech v3 Bland AI | 0.9 | — | — | — | 303 ms |
| 39 | Hume AI Octave 2 Hume AI | 0.6 | — | — | — | 448 ms |
| 40 | Alibaba Qwen3 TTS Flash Alibaba (Qwen) | 0.5 | — | — | — | 472 ms |
| 41 | Google Gemini 3.8 Flash-Lite TTS Google | 0.4 | — | — | — | 497 ms |
| 42 | ElevenLabs Eleven v4 ElevenLabs | 0.2 | — | — | — | 844 ms |
| 43 | Google Gemini 3.1 Flash TTS Google | 0.1 | — | — | — | 978 ms |
| 44 | Google Gemini 3.8 Flash TTS Google | 0 | — | — | — | 1031 ms |
Frequently asked questions
- Which TTS API should be used for real-time applications?
- On measured speed alone, Gradium TTS Beta ranks first: 58.7 of 60 speed points, with a P50 time-to-first-audio of 47.2 ms and a P99 of 108.8 ms on Coval's 30-day window ending October 1, 2026. Speed is not the whole choice: the overall ranking also weighs voice quality, accuracy and price.
- Which TTS API has the lowest P99 latency?
- Gradium TTS Beta: 108.8 ms P99 time-to-first-audio on Coval's 30-day window. The P99 is the slowest 1% of turns, the ones a caller notices.
- How is real-time speed measured here?
- With two independent benchmarks. Coval measures time-to-first-audio continuously over 30 days (thousands of runs per model): we use its mean, P99 and standard deviation. Speko publishes a P50 measured from US East (n=30 per model). Each source weighs by the square root of the number of runs it publishes, so Coval carries most of the speed score.
- Why are some TTS APIs missing from this page?
- This page lists only the 44 models whose time-to-first-audio Coval or Speko measure. The 68 others are in the overall ranking, where their speed counts zero and their page says why.