Speed · read on October 27, 2026
Fastest text-to-speech API: 31 TTS models in three speed tiers
For streaming and real-time TTS, latency is the first filter. These tables use the P95 time to first audio that Coval measures for each hosted API over 30 days from US East, grouped into three tiers. Models are listed in the order of the full ranking, fastest first.
Headroom: 4 models
P95 under 120 ms. The voice starts so fast that there is time left to spend elsewhere in the turn: speech recognition, the language model, a tool call.
| Rank | Model | Speed tier | P95 | Speed (shorter is faster) | Median | Listener Elo | Price per 1M chars | Best for |
|---|---|---|---|---|---|---|---|---|
| 1 | | Headroom | 75 ms | 39.7 ms | Not rated | $9 | Voice agents | |
| 2 |  Gradium TTSGradium | Headroom | 104.3 ms | 53.7 ms | 1,174 | $36 to $57.8 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 3 |  Inworld Realtime TTS-2 FlashInworld | Headroom | 100.9 ms | 58.1 ms | 1,218 | $10.4 to $15 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 4 |  Alibaba Qwen3 TTS FastAlibaba (Qwen) | Headroom | 98.8 ms | 62.9 ms | Not rated | $10 | Voice agents |
Fast enough: 9 models
P95 from 120 ms to 293 ms. Fast enough for most agents: the pause before the voice is short next to the rest of a turn.
| Rank | Model | Speed tier | P95 | Speed (shorter is faster) | Median | Listener Elo | Price per 1M chars | Best for |
|---|---|---|---|---|---|---|---|---|
| 5 |  Alibaba Qwen3 TTS 1.7bAlibaba (Qwen) | Fast enough | 136.6 ms | 103.6 ms | Not rated | Not collected | Voice agents | |
| 6 |  SpeechifyAI Simba 3.0Speechify | Fast enough | 149.4 ms | 100.5 ms | 1,116 | $6.6 to $10 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 7 |  SpeechifyAI Simba 3.2Speechify | Fast enough | 161.1 ms | 104.3 ms | 1,249 | $6.6 to $10 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 8 |  Inworld Realtime TTS-2Inworld | Fast enough | 194.7 ms | 149.1 ms | 1,267 | $15 to $25 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 9 |  Palabra TTS v1Palabra | Fast enough | 214.3 ms | 105.1 ms | Not rated | $12 | Voice agents | |
| 10 |  Fish Audio Drama 3 PreviewFish Audio | Fast enough | 209.3 ms | 146 ms | 1,245 | $15 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 11 |  ElevenLabs Eleven v4 TurboElevenLabs | Fast enough | 238.6 ms | 188.9 ms | 1,333 | $40 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 12 |  ElevenLabs Flash v2.5ElevenLabs | Fast enough | 255.1 ms | 191 ms | 1,076 | $40 to $50 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing | |
| 13 |  Deepgram Flux TTSDeepgram | Fast enough | 267 ms | 197.7 ms | 1,052 | $45 | Customer service, Voice agents, Assistants, Entertainment, Knowledge sharing |
Visible delay: 18 models
P95 above 293 ms. The voice becomes a visible part of the delay: a caller hears the wait.
| Rank | Model | Speed tier | P95 | Speed (shorter is faster) | Median | Listener Elo | Price per 1M chars | Best for |
|---|---|---|---|---|---|---|---|---|
| 14 |  Soniox TTS RT v1Soniox | Visible delay | 319 ms | 258.3 ms | Not rated | $13 to $14.1 | Not rated | |
| 15 |  Soniox TTS Real-Time v2Soniox | Visible delay | 324.7 ms | 264.4 ms | 1,180 | $13 to $14.81 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 16 |  Murf AI Falcon 2Murf AI | Visible delay | 363.5 ms | 171.3 ms | 1,154 | $10 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 17 |  Rime Mist v3Rime | Visible delay | 359.9 ms | 275.2 ms | Not rated | $30 | Not rated | |
| 18 |  Cartesia Sonic 3.6Cartesia | Visible delay | 374.4 ms | 285.8 ms | 1,292 | $39.2 to $50 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 19 |  Cartesia Sonic 3.5Cartesia | Visible delay | 390.5 ms | 279.3 ms | 1,190 | $39.2 to $50 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 20 |  Deepdub PhantomZ 4.0Deepdub | Visible delay | 430 ms | 183.4 ms | Not rated | Not collected | Not rated | |
| 21 |  Rime CodaRime | Visible delay | 428.1 ms | 317.5 ms | 1,054 | $50 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 22 | Out of Set Airy TTSOut of Set | Visible delay | 453.9 ms | 212.3 ms | Not rated | $3.67 | Not rated | |
| 23 |  SpaceXAI Grok TTSxAI | Visible delay | 473.9 ms | 329.8 ms | 1,155 | $15 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 24 |  Deepgram Aura 2Deepgram | Visible delay | 525.4 ms | 263 ms | Not rated | $30 | Not rated | |
| 25 |  Fish Audio S2.1 ProFish Audio | Visible delay | 533 ms | 328.1 ms | 1,184 | $15 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 26 | Guava Daytona TTSGuava | Visible delay | 584.8 ms | 450 ms | Not rated | Not collected | Not rated | |
| 27 |  Smallest.ai Lightning V3.1 Pro (Jul 2026)Smallest.ai | Visible delay | 703.3 ms | 360.5 ms | 1,169 | $19.5 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 28 |  Alibaba Qwen3 TTS Flash RealtimeAlibaba (Qwen) | Visible delay | 785.8 ms | 726.8 ms | Not rated | $13 | Not rated | |
| 29 |  Google Chirp 3: HDGoogle | Visible delay | 917.2 ms | 534.1 ms | 1,061 | $30 | Customer service, Assistants, Entertainment, Knowledge sharing | |
| 30 |  Deepgram Aura 2 EnDeepgram | Visible delay | 985.7 ms | 527.2 ms | Not rated | Not collected | Not rated | |
| 31 |  Fish Audio S2.1 Pro FreeFish Audio | Visible delay | 1,137.9 ms | 793.8 ms | Not rated | Not collected | Not rated |
Frequently asked questions
- What is the fastest text-to-speech API?
- By P95 time to first audio over Coval's 30-day window, the fastest 5 are Fluxions vui (75 ms), Gradium TTS (104.3 ms), Inworld Realtime TTS-2 Flash (100.9 ms), Alibaba Qwen3 TTS Fast (98.8 ms) and Alibaba Qwen3 TTS 1.7b (136.6 ms), in that order. The order among the leaders can change within a week, so each page shows the date of the reading.
- What is time to first audio?
- The delay between sending text to a TTS API and receiving the first piece of audio. It is the wait a caller hears before the voice starts. Coval publishes the median (P50) and the 95th percentile (P95), which is the slow turn.
- How are the speed tiers drawn?
- Each boundary is the middle of the widest gap between two consecutive P95 values, searched between 80 and 160 ms for the first and between 200 and 400 ms for the second, so no single model decides it. The current boundaries are 120 ms and 293 ms, recomputed at each reading.
- Why are only 31 of 96 models in a tier?
- Only Coval times models over 30 days, and it times 31 of them. The other 65 stream audio according to their vendor's documentation or another benchmark, but have no speed tier; nothing is estimated for them.