What is Alania and Sense? Turkish Voice Models of Patientdesk.ai
Alania and Sense are Patientdesk.ai's AI voice models trained for Turkish. What they do, why they matter, and who can use them.
Alania and Sense are two AI voice models that Patientdesk.ai trained from scratch for Turkish. Alania translates text to sound (TTS) and Sense translates speech to text (STT), both used with a single account and a single API key; the endpoint is compatible with the OpenAI API format, meaning it can be plugged into a project using Whisper or another audio service with minor changes.
Alania's highlight is the delay. It can start audio streaming before the sentence is complete; the company gives the first audio track delay it measures over the public API as a median of 190 milliseconds, including network time. The output is 48 kHz. This is exactly the measure that is decisive in delay-sensitive uses such as phone calls and voice assistants.
There are 11 ready-made sounds on the sound options side, male and female. The default female voice is F1, male voice is Öncel; others include Nisa, Emre, Selin and Melisa. Limit of 5,000 characters per request — longer text spins as a single audio stream, divided by sentence boundaries.
Two more features remain in the preview phase: describing the style of speech with text (slow and calm, whisper, cheerful, etc.) and describing a new sound with a description of age, tone, and energy without ever recording. The company itself states that style instructions are not always strictly enforced. Voice cloning is beta: it is a short recording and works with the text of that recording, the approval field must be sent in the request and permission must be obtained from the owner of the voice.
DUYU, on the other hand, transcribes both recorded audio files and live speech in real time. It gives a time stamp on a word basis; it can be printed in SRT and VTT subtitle formats.
The accuracy figures are as follows: In the open Turkish reading test FLEURS, the word error rate of Sense is 4.71%, and in the same test, OpenAI Whisper large-v3 is 5.04%. In the challenging test set prepared by Patientdesk itself — phone-quality voice, voice-over e-mail addresses and phone numbers — Sense gives 9.87% and ElevenLabs Scribe v2 gives 10.4%. All of these measurements are the results shared by the company itself; not an independent verification.
Here's why these numbers are important: most audio technologies are trained predominantly on English data, and Turkish is struggling with these models. Additive language structure, proper names, low-quality phone voice, and spoken e-mail or phone information are classic sources of error. A model specially trained for Turkish makes a difference at these points.
There is an introductory period on the pricing side: one month free from the date of registration, no card requested. The pay-as-you-go model applies — Alania is $0.015 per 1,000 characters, Sensory file processing is $0.125 per audio hour, and Sensory is $0.225 per live audio hour. Monthly billing and prices exclude VAT. On the corporate side, special quota, installation and data processing contract options are offered on your own server.
The company also took a step forward on the open-source side: it published its experimental Turkish TTS model, Antalia, on Hugging Face — along with model weights, code, technical report, and an approximate 5-hour Turkish speaking dataset. Patientdesk positions this as a separate research project from its commercial products. On the API side, Antalia-2 Mini always remains free.
The experience behind the models comes from an interesting place: the company says it has trained its Turkish models from the experience of a voice assistant who answers the phones of clinics in the US. Patientdesk.ai received a $1 million pre-seed investment in February 2026 under the leadership of Y Combinator.
Who can use it? Anyone who develops a Turkish voice assistant, call center automation, transcription and subtitling, voice memo, or accessibility product. To try it, simply register through speech.patientdesk.ai and get a single API key; you can also try the models through the browser.
In terms of the Turkish entrepreneurial ecosystem, the main note is that Turkish is becoming an area that is optimized on its own, rather than being a by-product of large models. This is a concrete example that shows that a language-trained model can pass global giants in their own test sets.
Note: This post was published on September 26, 2026; updated on November 5, 2026 with audio count, latency measurement, and current pricing information for Alania-2.
· 2026-09-26