A Comprehensive Polish Medical Speech Dataset for Enhancing Automatic Medical Dictation.

Area of Science:

• Speech Technology
• Medical Informatics
• Computational Linguistics

Background:

• Pre-trained models offer strong zero-shot capabilities but specialized domains like medical speech recognition require tailored datasets.
• Existing resources may not adequately capture the nuances of medical vocabulary and real-world acoustic conditions.

Purpose of the Study:

• To introduce ADMEDVOICE, a novel Polish medical speech dataset designed for improving medical speech recognition.
• To provide a comprehensive resource including original recordings, anonymized data, and synthetic data to facilitate research and development.

Main Methods:

• Collected nearly 15 hours of Polish medical speech data from 28 speakers under diverse, including noisy, conditions.
• Created anonymized and synthetic (text-to-speech) versions, expanding the dataset to over 83 hours and ~50,000 samples.
• Evaluated the Whisper model, fine-tuning it with the ADMEDVOICE dataset and its enhanced versions.

Main Results:

• The baseline Whisper model achieved a 24.03% Word Error Rate (WER) on the test set.
• Fine-tuning with human recordings reduced WER to 15.47%.
• Incorporating anonymized and synthetic data further decreased WER to 13.91%.

Conclusions:

• The ADMEDVOICE dataset significantly enhances medical speech recognition performance.
• Open-sourcing the dataset, fine-tuned model, and code promotes further advancements in the field.
• The combination of real, anonymized, and synthetic data offers a robust approach for model training.