A survey of automatic speech recognition deep models performance for Polish medical terms

Among the numerous applications of speech-to-text technology is the support of documentation created by medical personnel. There are many available speech recognition systems for doctors. Their effectiveness in languages such as Polish should be verified. In connection with our project in this field, we decided to check how well the popular speech recognition systems work, employing models trained for the general Polish language. For this purpose, we selected 100 words from the International Classification of Diseases dictionary, the Polish-language version of the International Statistical Classification of Diseases and Health Problems. The words were read into a microphone by five women and five men and also generated with a speech synthesizer using a male and a female voice. This resulted in 1,200 recordings tested with the following systems: Whisper, Google speech-to-text, and Microsoft Azure speech-to-text. The achieved word recognition performance is reflected by the calculated metrics: WER, WIL, Levenshtein distance, Jaccard distance, MER, and CER. Results show that the highest efficiency for most cases was obtained by Azure speech-to-text. However, none of the tested models is ready for voice-filling medical records, describing cases, or prescribing treatment, because the number of errors made when converting speech to text is too high.