Text to Speech Providers for a Voice Agent
The text to speech reads the answers of the agent while they are written, so the
user hears the first sentence before the last one exists. It is passed to the
server as tts.
Hosted
- ElevenLabs, high-quality voice synthesis, with audio tags on Eleven v4 Turbo
- Cartesia, low-latency streaming TTS
- Gemini TTS, 30 voices, directed with plain instructions
- Mistral Voxtral Speech, Voxtral Speech 3 in 35+ languages, with voice cloning
- OpenAI,
gpt-realtime-2.1-minior the speech endpoint models - Gradium, streaming voices on an EU endpoint
Local
- Kokoro, in your Node process, English only
- Phonon, by Gradium, in your Node process, five languages
- Piper, 53 languages
- Pocket TTS, voice cloning in your Node process, English only
- Qwen3-TTS, ten languages, on an mlx-audio server
The local text to speech guide helps pick a model.
Comparison
| Provider | Latency | Quality | Voices | Cost per hour | Best For |
|---|---|---|---|---|---|
| ElevenLabs | ~400ms | Excellent | 1000+ | $1.80 (Flash v2.5), $3.60 (Multilingual v2) | High-quality voices |
| ElevenLabs v4 Turbo | ~100ms | Excellent | 1000+ | $1.80 | Expressive voices, audio tags |
| Cartesia | ~150ms | Good | 50+ | $1.80 | Low-latency streaming |
| Gradium 🇫🇷 | ~300ms | Excellent | 150+ | $2.60 | French, EU data residency |
| Kokoro (local) | ~1000ms | Good | 28 | Free | Privacy, English only |
| Phonon (local) 🇫🇷 | ~50ms | Very good | 102 | Free | Privacy, five languages |
| Piper (local) | ~390ms | Fair | 178 | Free | Privacy, 53 languages |
| Qwen3-TTS (local) | ~400ms | Very good | 9 | Free | Privacy, ten languages |
Costs are the published list price of the entry paid plan in August 2026, per hour of audio generated. ElevenLabs bills its API in dollars at the same price on every plan, $40 per million characters on Eleven v4 Turbo and Flash v2.5 and $80 on Multilingual v2 (list prices read on October 20, 2026). Every provider charges less as the volume grows. The prices per character are converted at the 45,000 characters an hour Gradium publishes for its own voices. The voice AI blog ranks more providers on measured latency and price.
Your own
Any other service plugs in by extending the TTS class, and
FallbackTTS switches to another provider when one fails.