The development of an automatic speech recognition model using interview data from long-term care for older adults

Area of Science:

• Gerontology
• Speech Technology
• Health Informatics

Background:

• Interviews in long-term care (LTC) are vital for understanding older adults' perspectives.
• Manual transcription of these interviews is time-consuming and labor-intensive.
• Existing automatic speech recognition (ASR) systems often perform poorly with specific demographics, including older adults and accented speech.

Purpose of the Study:

• To develop an effective ASR model tailored for interview data from older adults in LTC settings.
• To demonstrate the efficacy of using specific demographic data to improve ASR performance.
• To reduce the word error rate (WER) and time required for transcribing qualitative data in LTC research.

Main Methods:

• An initial ASR model was created using the Mozilla Common Voice dataset.
• The model was fine-tuned using 34 hours of audio and transcript data from interviews with LTC residents, families, and staff.
• Continuous processing and refinement of interview data were employed to minimize the word error rate (WER).

Main Results:

• The initial ASR model had a high WER of 48.3% on interview data due to background noise and mispronunciations.
• Fine-tuning with interview data reduced the average WER to 24.3%, with a median WER of 22.1% on test data.
• The improved ASR model was over six times faster than manual transcription, with residents' speech showing the highest WER at 22.7%.

Conclusions:

• Fine-tuning ASR models with specific interview data significantly decreases the word error rate, proving the method's effectiveness.
• The developed ASR system offers a substantial time-saving solution for transcribing qualitative data in long-term care research.
• Local transcription using the improved ASR model can enhance participant privacy by minimizing data sharing.