Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Streaming TTS APIs in 2026: which multilingual models start speaking before the text is done
Streaming TTS starts producing audio before the whole sentence is available, which is what a voice agent needs when a language model writes the answer word by word. All 23 ranked models are marked streaming; 12 also have a Cekura result for accepting streamed text input.
Cekura reports the median time to first audio with streamed text and with the full text; the two columns show how much streaming input helps each model.
How the rows are ordered. Same order as the ranking: speed tier, then P95, ties by median.
| Rank | Model | Streaming Measured by a benchmark, or described in vendor documentation | Accepts streamed text | First audio, streamed text (ms) Cekura, median | First audio, full text (ms) Cekura, median | P95 (ms) 1 turn in 20 is slower |
|---|---|---|---|---|---|---|
| 1 = |  Gradium TTS | Measured | Yes | -64 | 76 | 104.3 |
| 1 = |  Inworld Realtime TTS-2 Flash | Measured | Yes | 45 | 46 | 100.9 |
| 1 = |  Alibaba Qwen3 TTS Fast | Measured | n/a | n/a | n/a | 98.8 |
| 4 |  Alibaba Qwen3 TTS 1.7b | Measured | n/a | n/a | n/a | 136.6 |
| 5 |  SpeechifyAI Simba 3.0 | Measured | n/a | n/a | n/a | 149.4 |
| 6 |  Inworld Realtime TTS-2 | Measured | Yes | 127 | 126 | 194.7 |
| 7 |  Palabra TTS v1 | Measured | n/a | n/a | n/a | 214.3 |
| 8 |  ElevenLabs Eleven v4 Turbo | Measured | Yes | 69 | 176 | 238.6 |
| 9 |  ElevenLabs Flash v2.5 | Measured | Yes | 177 | 182 | 255.1 |
| 10 = |  Soniox TTS RT v1 | Measured | n/a | n/a | n/a | 319.0 |
| 10 = |  Soniox TTS Real-Time v2 | Measured | Yes | 183 | 241 | 324.7 |
| 12 = |  Murf AI Falcon 2 | Measured | n/a | n/a | n/a | 363.5 |
| 12 = |  Rime Mist v3 | Measured | Yes | 293 | 292 | 359.9 |
| 14 |  Cartesia Sonic 3.6 | Measured | Yes | 270 | 286 | 374.4 |
| 15 |  Cartesia Sonic 3.5 | Measured | Yes | 244 | 261 | 390.5 |
| 16 = |  Deepdub PhantomZ 4.0 | Measured | n/a | n/a | n/a | 430.0 |
| 16 = |  Rime Coda | Measured | Yes | 269 | 282 | 428.1 |
| 18 |  SpaceXAI Grok TTS | Measured | Yes | 157 | 358 | 473.9 |
| 19 |  Fish Audio S2.1 Pro | Measured | Yes | 321 | 333 | 533.0 |
| 20 |  Smallest.ai Lightning V3.1 Pro (Jul 2026) | Measured | n/a | n/a | n/a | 703.3 |
| 21 |  Alibaba Qwen3 TTS Flash Realtime | Measured | n/a | n/a | n/a | 785.8 |
| 22 |  Google Chirp 3: HD | Measured | n/a | n/a | n/a | 917.2 |
| 23 |  Fish Audio S2.1 Pro Free | Measured | n/a | n/a | n/a | 1,137.9 |
Frequently asked questions
- What is a streaming TTS API?
- A text-to-speech API that returns audio in chunks while synthesis is still running, and in the best case also accepts text in chunks. It lets a voice agent speak before the full reply exists.
- Which TTS APIs accept streamed text input?
- Starting with the best-ranked, Gradium TTS, Inworld Realtime TTS-2 Flash, Inworld Realtime TTS-2, and ElevenLabs Eleven v4 Turbo have a Cekura result for streamed text input. Models without a result are marked n/a, not no.
- How is streaming status decided?
- Streaming is marked Measured when a benchmark times the model's first audio, and Vendor documentation when only the vendor's streaming API page describes it.
- Does streaming reduce latency?
- Cekura's two columns show it per model.