A Small‑Clinic Friendly Speech‑to‑Text Model for Doctor‑Patient Conversations
CobuildX‑AI released a LoRA‑adapted speech‑to‑text model that improves medical terminology accuracy while running on a single GPU or CPU.

The cbx‑healthcare‑speech‑to‑text‑0.6B model boosts condition‑name accuracy to 94.6% and medicine‑name accuracy to 66.1% compared with its base. It is small enough for a typical clinic computer and can batch recordings for faster turnaround.
Why a medical‑focused ASR matters for small practices
Most off‑the‑shelf speech‑to‑text services are trained on general conversation and often mis‑recognise clinical terms. For a small clinic that relies on quick documentation, those errors translate into extra time spent correcting notes or, worse, missing critical details. A model tuned specifically on doctor‑patient exchanges can reduce that friction.
The CobuildX‑AI model is built on the Qwen3‑ASR‑0.6B base and adds a LoRA adapter trained on 28 hours of real and synthetic medical consultations. According to its evaluation, it identifies medical condition names correctly 94.6% of the time—up from 83.9% for the base model—and gets medicine names right 66.1% of the time, compared with 61.2% before adaptation.
Performance numbers you can rely on
In testing on the PriMock57 dataset—57 mock UK GP consultations that were never part of training—the model achieved a word error rate of 6.01%, essentially matching the base model for everyday words. The biggest gains appear in clinical vocabulary: condition‑name mistakes dropped from 16.1% to 5.4%, and medicine‑name mistakes fell from 38.8% to 33.9%. Common medicines such as paracetamol and ibuprofen were almost always correct.
Speed is also practical for a clinic setting. On an Apple M4 CPU, one minute of speech processes in about 10–15 seconds; on an NVIDIA T4 GPU the same minute takes roughly 2.4 seconds. A full 10‑minute consultation therefore finishes in about two minutes on CPU or 25 seconds on GPU, allowing near‑real‑time turnaround for batch uploads.
Running the model on modest hardware
The model can run on a single GPU or even a standard CPU. When using a GPU, it operates in fp16 precision and can handle batches of 16 clips, which speeds up bulk transcription. On CPU it runs in fp32 with four threads, which is sufficient for a small practice that processes a few recordings each day.
Long recordings are automatically split into 30‑second segments at quiet moments, then re‑assembled into a single transcript. The output retains capitalization, punctuation, and numbers, while filler sounds like “um” or “uh” are omitted. This reduces the amount of manual cleanup needed before the text can be added to patient records.
Limitations to keep in mind
The model only supports English and was trained on US‑based accents. Accuracy on other accents has not been fully measured, so clinics with diverse patient populations should validate performance on their own recordings.
Medicine‑name accuracy, while improved, still misses about one in three mentions, especially for rare or UK‑specific brand names such as Clenil, Implanon, and Dioralyte. Human review of transcripts remains essential for clinical documentation.
Because the model is released under an Apache‑2.0 license, you can integrate it into internal tools without additional fees, but you must ensure any downstream use complies with the license terms.
Getting started without a big investment
To try the model, download it from Hugging Face and install the listed Python requirements. The repository includes a simple wrapper that loads the base Qwen3‑ASR‑0.6B model together with the LoRA adapter. Once loaded, a single function call can transcribe a WAV file or a batch of files, optionally providing a context list of expected medication names to improve recognition.
For a clinic that already records consultations, the workflow can be as simple as: (1) place the audio files in a folder, (2) run the transcription script, and (3) import the resulting JSON into the electronic health record system. Because the code runs locally, patient data never leaves the practice, addressing privacy concerns that arise with cloud‑based ASR services.
Takeaways
- A domain‑specific ASR model can dramatically improve the accuracy of medical terms while keeping hardware requirements low.
- Even on a standard CPU, the model processes a 10‑minute consultation in about two minutes, making batch transcription feasible for small clinics.
- Human review is still needed for rare medication names and non‑US accents, but the tool reduces the manual typing workload.
Where would this work in your business?
Take the two-minute AI Opportunity Snapshot and we will point to where automation is most likely to pay off for you.