Patientdesk Labs is the research arm of Patientdesk.ai. We build benchmarks, fine-tune models, and develop domain-specific tools so that when an AI answers the phone at a dental office, it actually works.
Dental clinic AI has unique requirements that generic models don't meet. Our research focuses on four areas where domain-specific work makes the biggest difference — from the model that reasons, to the voice the patient actually hears.
The first benchmark for evaluating LLMs as dental clinic phone agents. 483 scenarios across 10 categories, scoring empathy, clinical safety, accuracy, brevity, and tone — plus a deployment-weighted leaderboard.
Read the paper →Soul-document-driven fine-tuning of Gemma 4 31B with Opus 4.6 as judge. After one iteration of SFT + DPO: 8.46 on DentesBench — beating every frontier API model including Opus itself.
Read the paper →A blind listening leaderboard for AI voices. Votes come from a vetted panel of paid native speakers rather than whoever finds the page, and each listener answers one of four questions instead of a single “which is better”. Across 12,245 comparisons, Fish Audio’s s2-pro is 5th of 44 on sounds human and 40th on clear and correct — and Cartesia’s sonic-3.5 is 3rd in English, 19th in Turkish.
Open the arena →Adapting speech-to-text models for dental clinic phone audio. Patient calls with accents, background noise, and dental terminology that generic models consistently get wrong — "prophylaxis" shouldn't become "prophy lax is."
Paper coming soonA dental receptionist AI has to be warm without accidentally diagnosing, efficient without being cold, and helpful without overstepping clinical boundaries. No off-the-shelf model gets this right consistently.
A vetted panel of paid native speakers listens to two AI voices reading the same sentence and picks one, without ever seeing who made either. Each sitting asks one of four questions and never mixes them: overall preference, sounds human, clear and correct, rhythm and expression. 12,245 comparisons across 44 English and 28 Turkish voices from 16 companies, judged by 224 listeners and frozen on 25 July. The two boards below both show overall preference, one language each — note how little they share. Open the full arena →
With 4.6× the data of our pilot, English now separates: 20 of the 43 challengers sit clearly behind the leader. The language gap is still the biggest single finding — Cartesia’s sonic-3.5 is 3rd in English and 19th in Turkish, ElevenLabs’ eleven_v3 is 37th and 8th, Google’s Gemini 2.5 Puck 38th and 4th; Gemini 3.1 Flash Puck, top of both boards above, is the exception. A vetted panel is not a representative one, and it is not infallible either: 9.3% of the hidden same-clip controls still got a confident winner picked.
Eight models evaluated on 483 dental phone agent scenarios. The v2 score weights quality (80%), cost (10%), and latency (10%) to reflect real deployment constraints. Full methodology →
Read the full DentesBench paper for methodology, results, and what we've learned about the tradeoffs in dental AI.
Read the Paper Visit Patientdesk.ai