Evaluating the BPPT Medical Speech Corpus for An ASR Medical Record Transcription System
Elvira Nurfadhilah, Asril Jarin, Lyla Ruslana Aini, Siska Pebiana, Agung Santosa, Mohammad Teduh Uliniansyah, Eduward Butarbutar, Desiani, Gunarso · 2021
In a joint research funded by Indonesia Endowment Fund for Education (LPDP), Agency for the Assessment and Application of Technology (BPPT) is collaborating with Solusi247 and Harapan Kita National Heart and Vascular Hospital to develop a speech corpus named BPPT Medical Speech Corpus for building a medical ASR (automatic speech recognition) system. The corpus was produced by recording 100 speakers for 81.68 hours. The corpus has 600 unique sentences containing medical terms. Previously, BPPT has developed a speech corpus for the general named BPPT General Speech Corpus. A third speech corpus named BPPT Combined Speech Corpus was created by combining the medical and general corpus, and each corpus was trained using PyChain to build an ASR model. Based on the experiments we conducted, the ASR model built from the combined speech corpus has the best WER of 6.10%.