Indian patients almost always say a medicine by its brand name, in English, no matter which language they are speaking. A multilingual speech-to-text model hears that brand name and aggressively “corrects” it into the nearest real English words it knows. The result is grammatically fine and clinically meaningless:
- “Azithral” becomes as it roll
- “Combiflam” becomes come be flame
- “Fycare-D” becomes fat cared
For a voice agent in healthcare, that single substitution can quietly break the whole record. IndMedSTT-Synth is our first public dataset built to study and fix exactly this failure.
What is inside
Each row pairs a true medication brand name with the medical context it is prescribed for, a naturalistic patient sentence as it would arrive from speech-to-text, and the precise misheard phrase that replaced the drug name. It is designed for post-ASR medication entity recognition and for training error-correction on Indian, code-mixed patient speech.
Version 1.0 holds 87 manually reviewed samples, drawn from a public Indian pharmaceutical database and generated with GPT-4o to simulate realistic autocorrection behaviour. Samples that were clinically implausible or linguistically unrealistic were corrected or dropped. A companion voice dataset, IndMedSTT-Voice, is in preparation.
Why we are releasing it
We publish what we learn. Getting medication capture right is foundational to everything IndRAH builds, and the failure mode above is under-studied for Indian multilingual speech. The full dataset, schema, and citation live on Hugging Face.