Freya TTS
Introducing Freya TTS Adam V1
The most realistic text-to-speech model, by blind listener vote. Second only to human speech on Audio Realism Bench, ahead of every other model.
Adam V1 ships with two English voices, Finn and Nina.
Freya TTS
Hear it.
Pick a scene. Edit the text if you like, then generate.
Audio Realism Bench
Blind listeners rank it second only to humans.
Design Arena's Audio Realism Bench plays two clips to native English listeners, who pick the one that sounds more human. Scores are Elo ratings fitted with Bradley-Terry. Freya was measured on 16 September 2026. The other rows are the public board on the same day.
- 1Human speech1462
- 2Freya TTS Adam V11419
- 3Bland Speech v31369
- 4Kalpa TTS Beta V0.11343
- 5Cartesia Sonic 3.61306
- 6MAI-Voice-21213
- 7Eleven v3 Conversational1202
- 8Cartesia Sonic 3.51188
- 9Grok TTS1150
- 10Murf Falcon 21141
- 11MiniMax Speech 2.8 HD1130
- 12Google Gemini 2.5 Pro TTS1101
- 13MiniMax Speech-02 HD1101
- 14OpenAI GPT Realtime 21071
- 15Google Gemini 3.1 Flash TTS1034
- 16OpenAI GPT-4o Mini TTS1022
- 17OpenAI GPT Realtime 2.11017
- 18Google Gemini 2.5 Flash TTS980
- 19Eleven v3964
- 20Lightning v3.1 Pro947
Latency
122 ms to first audio.
Audio starts streaming before a person would have begun to reply.
Measured to the first audio chunk on the streaming endpoint, on the same machine as the model, the way it runs in an on-prem deployment. The human figure is the typical gap before a reply in conversation, about 200 ms (Stivers et al., 2009).
Run it in your own data center.
Adam V1 ships as an API and as an on-prem package for banks and insurers. Turkish voices come from FreyaTTS-small.