Freya Research2026
Turkish speech synthesis
FreyaTTS-small
Our compact model, trained from scratch to turn Turkish text into speech. Read the paper, inspect the results and run it on your own infrastructure.
- Audio output
- 48 kHz
- License
- Apache 2.0
BLIND ORDER
Which one is human?
Tap a flame to hear each voice. Hear both, then choose.
A voice model so human, the only honest benchmark is your ear.
ADAM V1 / ENGLISH VOICE MODEL / 2026
And then,
it spoke.
0 / 6
I / NEWS · DESIGN ARENA, AUDIO REALISM BENCH
We built the world’s most realistic voice model.
Freya TTS Adam V1 ranks first among models on Design Arena’s blind listening test, just behind human speech. Listeners hear two clips and pick the one that sounds more human. The ranking is an Elo rating.
Audio Realism Bench
DesignArenaby Intelligence
DesignArenaby Intelligence- 1462Human speech
- 1419Freya TTS Adam V1
- 1369Bland Speech v3
- 1306Cartesia Sonic 3.6
- 1202Eleven v3 Conversational
- 1188Cartesia Sonic 3.5
- 1150Grok TTS
- 1101Google Gemini 2.5 Pro TTS
- 1071OpenAI GPT Realtime 2
II / THE API
Build with
the voice.
Two English voices, Adam and Eve, behind one API. Access is by request while we open it in batches.
III / FREYA RESEARCH
We build our own speech models.
We work on speech synthesis, speech recognition and voice interaction. With FreyaTTS-small, we have released our code, model weights and evaluation dataset for anyone to use.
Explore the researchHelp build the next model.
Working on speech or language models? We'd like to meet you. Email us your research, code or the questions you want to explore.
We also support emerging researchers. Discover the YZ50 program