Patientdesk.ai has launched two voice models designed for the Turkish language: Alania for text-to-speech and Duyu for speech-to-text. Both models can be used through a single account and API key, with a one-month free usage period from the registration date as part of the launch offer.
Models Designed for the Challenges of the Turkish Language
The company says development of the two models began with the problem that many voice-AI solutions rely primarily on English-language data. It adds that the agglutinative nature of Turkish, proper nouns, reduced audio quality in telephone calls, as well as the pronunciation of email addresses and phone numbers, present practical challenges for general-purpose models.
Patientdesk.ai drew on its experience developing a voice assistant that answers calls for clinics in the United States, then trained Alania and Duyu with a focus on Turkish use cases.
Duyu’s Speech-to-Text Performance
According to figures published by the company, Duyu achieved a word error rate of 4.71% on the open Turkish FLEURS test, compared with 5.04% for OpenAI’s Whisper large-v3 model on the same test.
In an internal test set containing telephone-quality audio and conversations mentioning email addresses and phone numbers, Duyu recorded an error rate of 9.87%, while the company reported a rate of 10.4% for ElevenLabs Scribe v2 on the same set. The article does not explain the test set’s size or methodology, so the comparison should be read as company-reported results rather than a comprehensive independent evaluation.
Duyu can process recorded audio files and convert live conversations into text in real time, making it suitable for real-time transcription and voice-call applications.
Alania and Low-Latency Audio Streaming
Alania focuses on reading Turkish text in a consistent, natural-sounding voice. The model can send audio as a stream before a sentence is complete, a feature suited to telephone calls and voice assistants where response speed matters. The platform currently offers one voice option, with intonation options associated with contexts such as appointments, questions, and calming.
API Integration and an Open Research Project
Both models are compatible with the common OpenAI API format, allowing developers to integrate them into existing applications with limited modifications. Patientdesk.ai also allows users to try the models through a browser and create API keys.
In parallel, the company released Antalia-1 as an experimental open-source Turkish text-to-speech model. The release includes the model’s weights and code, a technical report, and approximately five hours of Turkish audio data. The company distinguishes this research project from its commercial products.
What Changes in Practice?
The launch gives developers an option aimed at specific Turkish use cases, particularly calls and data containing numbers and email addresses. However, the published comparisons alone are insufficient to assess quality across all environments, and the availability of only one voice in Alania currently limits usage flexibility. Antalia-1, meanwhile, gives researchers an inspectable and developable starting point, while performance evaluation and practical licensing remain outside the details provided in the article.