Patientdesk.ai, a Turkish healthcare AI startup that raised $1M in a Y Combinator-led pre-seed round earlier this year, has launched two voice models designed specifically for the Turkish language: Alania for text-to-speech and Duyu for speech-to-text.
Turkish poses unique challenges for voice AI: its agglutinative grammar, proper nouns embedded in phone calls, and low-quality telephone audio all trip up models trained primarily on English. Patientdesk.ai built both models from experience developing a phone-answering voice assistant for US clinics, then retrained specifically for Turkish.
On the FLEURS Turkish reading benchmark, Duyu achieves a 4.71% word error rate, outperforming OpenAI's Whisper large-v3 (5.04%) on the same test. In the company's own test set — composed of telephone-quality audio and spoken email addresses and phone numbers — Duyu scores 9.87% versus ElevenLabs Scribe v2's 10.4%. Duyu also supports real-time streaming transcription.
Alania produces natural Turkish speech with low latency, streaming output before the full sentence completes — a design choice aimed at phone-based assistants where delays matter. It currently offers a single voice with tone options for appointment scheduling, questions, and calming speech.
Both models use the OpenAI-compatible API format, allowing developers to swap them into existing applications with minimal changes. Patientdesk.ai also open-sourced Antalia-1, an experimental single-voice TTS model, along with its weights, code, technical report, and a 5-hour Turkish audio dataset. The commercial models are free for one month from registration.
