Text to speech that reads like a person talks.
Send text, get natural speech back while it is still being generated. One multilingual model covers 39 languages, and the newest, Kugel 3.5, gets numbers, names and one-word replies right far more often than the model before it.
Start free. No credit card.
First audio, measured.
Median time to first audio across leading streaming voice models.
WebSocket streaming. Network round trip included, connection setup excluded. Lower is better.
| Model | Median |
|---|---|
| Elbic Kugel 3 | 91 ms |
| Cartesia Sonic 3.5 | 111 ms |
| ElevenLabs Flash v2.5 | 126 ms |
| Cartesia Sonic Turbo | 127 ms |
| Deepgram Aura-1 | 157 ms |
| ElevenLabs Turbo v2.5 | 178 ms |
| Deepgram Aura-2 | 207 ms |
| ElevenLabs Multilingual v2 | 431 ms |
Tech specs
Read the docs- Models
- kugel-3.5, the default; kugel-3 when you set it
- Languages
- 39, from one multilingual model
- Time to first audio
- 91 ms median (Kugel 3)
- Streaming
- Text in, audio out, over REST or WebSocket
- Voice cloning
- From your own reference audio
- Audio output
- 24 kHz native; 8, 16, 22.05, 24 and 44.1 kHz. 8 kHz for telephony, no resampling step
- Input
- Up to 10,000 characters per request
- Control
- Break tags, spell tags, IPA, pronunciation dictionaries, speed control, word-level timestamps
- Hosting
- Built, trained and hosted in the EU
Everything is in the API.
Streaming both ways
Stream text in as your LLM writes it and flush once per turn. The server groups text at sentence boundaries and streams audio back.
Voice cloning
Create a voice from a reference recording. Clean audio of a single speaker works best.
Text normalization
Numbers, dates and currencies become words before synthesis, so you send text as you would write it.
Pronunciation
Write IPA between slashes, or keep a pronunciation dictionary for a whole project.
Make the first call today.
Create an account in a minute and start free, with no credit card. Or talk to us about on-premise hosting, SLAs and committed rates.