Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Best TTS APIs for voice agents in 2026: streaming, latency and customer-service ratings
A voice agent cannot wait for a whole sentence to be synthesized, so every model here streams audio. The speed leaders are Gradium TTS, Inworld Realtime TTS-2 Flash, and Alibaba Qwen3 TTS Fast; listener ratings in the customer-service category are in the column beside the latency.
Voice agent platforms such as Vapi, Retell, LiveKit and Twilio orchestrate speech-to-text, a language model and a TTS voice. They are not TTS models, so they are not ranked here: you pick a TTS API from this list and connect it through the platform.
How the rows are ordered. Same order as the ranking: speed tier, then P95, ties by median.
| Rank | Model | Speed tier From the Coval P95 | P95 (ms) 1 turn in 20 is slower | Streaming Measured by a benchmark, or described in vendor documentation | Accepts streamed text | Customer-service Elo Artificial Analysis, customer-service category | Rated for |
|---|---|---|---|---|---|---|---|
| 1 = |  Gradium TTS | Headroom | 104.3 | Measured | Yes | 1,241 ±26 | Knowledge sharing, Assistants, Entertainment, Customer service, Voice agents |
| 1 = |  Inworld Realtime TTS-2 Flash | Headroom | 100.9 | Measured | Yes | 1,244 ±25 | Knowledge sharing, Assistants, Entertainment, Customer service, Voice agents |
| 1 = |  Alibaba Qwen3 TTS Fast | Headroom | 98.8 | Measured | n/a | n/a | Voice agents |
| 4 |  Alibaba Qwen3 TTS 1.7b | Fast enough | 136.6 | Measured | n/a | n/a | Voice agents |
| 5 |  SpeechifyAI Simba 3.0 | Fast enough | 149.4 | Measured | n/a | 1,130 ±25 | Knowledge sharing, Assistants, Entertainment, Customer service, Voice agents |
| 6 |  Inworld Realtime TTS-2 | Fast enough | 194.7 | Measured | Yes | 1,311 ±28 | Knowledge sharing, Assistants, Entertainment, Customer service, Voice agents |
| 7 |  Palabra TTS v1 | Fast enough | 214.3 | Measured | n/a | n/a | Voice agents |
| 8 |  ElevenLabs Eleven v4 Turbo | Fast enough | 238.6 | Measured | Yes | 1,372 ±33 | Knowledge sharing, Assistants, Entertainment, Customer service, Voice agents |
| 9 |  ElevenLabs Flash v2.5 | Fast enough | 255.1 | Measured | Yes | 1,089 ±26 | Knowledge sharing, Assistants, Entertainment, Customer service, Voice agents |
| 10 = |  Soniox TTS RT v1 | Visible delay | 319.0 | Measured | n/a | n/a | n/a |
| 10 = |  Soniox TTS Real-Time v2 | Visible delay | 324.7 | Measured | Yes | 1,213 ±21 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 12 = |  Murf AI Falcon 2 | Visible delay | 363.5 | Measured | n/a | 1,179 ±25 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 12 = |  Rime Mist v3 | Visible delay | 359.9 | Measured | Yes | n/a | n/a |
| 14 |  Cartesia Sonic 3.6 | Visible delay | 374.4 | Measured | Yes | 1,331 ±28 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 15 |  Cartesia Sonic 3.5 | Visible delay | 390.5 | Measured | Yes | 1,203 ±24 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 16 = |  Deepdub PhantomZ 4.0 | Visible delay | 430.0 | Measured | n/a | n/a | n/a |
| 16 = |  Rime Coda | Visible delay | 428.1 | Measured | Yes | 1,108 ±26 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 18 |  SpaceXAI Grok TTS | Visible delay | 473.9 | Measured | Yes | 1,193 ±27 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 19 |  Fish Audio S2.1 Pro | Visible delay | 533.0 | Measured | Yes | 1,213 ±27 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 20 |  Smallest.ai Lightning V3.1 Pro (Jul 2026) | Visible delay | 703.3 | Measured | n/a | 1,189 ±24 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 21 |  Alibaba Qwen3 TTS Flash Realtime | Visible delay | 785.8 | Measured | n/a | n/a | n/a |
| 22 |  Google Chirp 3: HD | Visible delay | 917.2 | Measured | n/a | 1,108 ±27 | Knowledge sharing, Assistants, Entertainment, Customer service |
| 23 |  Fish Audio S2.1 Pro Free | Visible delay | 1,137.9 | Measured | n/a | n/a | n/a |
Frequently asked questions
- What should I look for in a TTS API for a voice agent?
- First the P95 latency, because slow turns are what callers notice; then whether the API accepts streamed text, so synthesis starts before the language model finishes; then voice quality in your use case. This table gives all three side by side without combining them.
- Which TTS APIs are rated for customer service?
- ElevenLabs Eleven v4 Turbo, Cartesia Sonic 3.6, and Inworld Realtime TTS-2 have the highest customer-service Elo among ranked models, starting with the highest. Artificial Analysis rates 15 ranked models for that category.
- Do I need a multilingual model for a voice agent?
- Only if callers speak more than one language. Every model here is tagged multilingual by Coval, but only the languages tested column shows evidence in a given language.
- Are Vapi, Retell or LiveKit TTS models?
- No. They are orchestration platforms that connect speech-to-text, a language model and a TTS voice. This site ranks the TTS models you can plug into them.