Freya

Freya is releasing Adam and Eve, two English text-to-speech voices aimed at developers and enterprises building voice agents, available through a single API with access rolling out in batches by request. The launch matters because Freya TTS Adam V1 ranks first among models on Design Arena's blind listening test, just behind human speech, where listeners hear two clips and pick the one that sounds more human , placing it above entries from ElevenLabs, Cartesia, Google, and OpenAI on the AudioRealismBench leaderboard. The pitch to founders and operators is a synthesis model close enough to natural speech that the team is willing to be benchmarked against a real human on the front page. The company was founded in 2025 by Tunga Bayrak, based in Istanbul, Turkey, and is part of Y Combinator 's S25 batch, with Tomas Nepala as co-founder. Beyond the consumer-facing demo, Freya's commercial focus is enterprise voice agents for call centers, particularly in financial institutions such as banks, fintechs, and insurers, where trust, accuracy, and data security matter most, spanning inbound support, outbound campaigns, and lead qualification. According to the launch video, the team spent 18 months training 27 voice models across autoregressive neural codecs, masked token transformers, and diffusion-based generators before settling on a flow-matching architecture with distillation for the shipping version. Freya has also open-sourced adjacent research, including FreyaTTS-small, a compact model trained from scratch to turn Turkish text into speech, released under Apache 2.0 with the paper, code, model weights, and evaluation dataset available , which gives outside researchers a way to inspect how the group approaches speech synthesis before committing to the closed API.