A Comprehensive Polish Medical Speech Dataset for Enhancing Automatic Medical Dictation

Pre-trained models have become widely adopted for their strong zero-shot performance, often minimizing the need for task-specific data. However, specialized domains like medical speech recognition still benefit from tailored datasets. We present ADMEDVOICE, a novel Polish medical speech dataset, collected using a high-quality text corpus and diverse recording conditions to reflect real-world scenarios. The dataset includes domain-specific vocabulary such as drug names and illnesses, with nearly 15 hours of audio from 28 speakers, including noisy environments. Additionally, we release two enhanced versions: one anonymized for privacy-sensitive use and another synthetic version created via text-to-speech, totaling over 83 hours and nearly 50,000 samples. Evaluating the Whisper model, we observe a 24.03 WER on our test set. Fine-tuning with human recordings reduces WER to 15.47, and incorporating anonymized and synthetic data further lowers it to 13.91. We open-source the dataset, fine-tuned model, and code on Kaggle to support continued research in medical speech recognition.