Best Text-to-Speech APIs for Voice Agents (2026)
A phone agent's voice has to start speaking fast and sound natural. The model table lists every TTS model we track with the latency, language count and price each vendor publishes. Most bill per 1,000 characters, which Rime equates to roughly one minute of speech; vendors measure latency differently, so treat those numbers as each vendor's own claim.
ElevenLabs, Cartesia, Rime, Deepgram, Speechmatics and Soniox TTS models compared on published latency, languages and price.
TTS models compared
| Model | Price | Latency | Languages | Cloning | |
|---|---|---|---|---|---|
| ElevenLabs Flash v2.5 | $0.04 / 1K chars | ~75 ms (model inference) | 32 | Yes | Source |
| ElevenLabs v4 Turbo | $0.04 / 1K chars | ~100 ms median (model inference) | 90+ | Yes | Source |
| Cartesia Sonic 3.6 | Free tier; from $5/mo (100K credits) | — | 44 | Yes (instant, paid plans) | Source |
| Rime Mist v3 | $0.03 / 1K chars (~$0.03/min) | 37 ms P50 time-to-first-audio | 4 (en, fr, de, es) | Enterprise | Source |
| Rime Coda | $0.05 / 1K chars (~$0.05/min) | 96 ms P50 time-to-first-audio | 8 | Enterprise | Source |
| Deepgram Aura-2 | $0.030 / 1K chars | — | 7 (en, es, de, fr, nl, it, ja) | — | Source |
| Deepgram Flux TTS | $0.045 / 1K chars | — | — | — | Source |
| Speechmatics TTS | $0.011 / 1K chars; first 1M chars free | — | — | — | Source |
| Soniox TTS | ~$0.70 / hr of speech | Real-time streaming | — | — | Source |
See all voice models · Latency and prices as each vendor publishes them.
ElevenLabs Agents vs Vapi →Deepgram vs AssemblyAI →Cartesia vs ElevenLabs →Deepgram vs Speechmatics →Rime vs Cartesia →Deepgram vs ElevenLabs →
Prices from vendors' official pricing pages, checked October 14, 2026. Not sponsored; order isn't a ranking.