Post-Processing Automatic Transcriptions with Machine Learning for Verbal Fluency Scoring.

Area of Science:

• Cognitive Psychology
• Computational Linguistics
• Speech Technology

Background:

• Verbal fluency tasks are crucial for assessing cognitive function.
• Accurate transcription with word timings is essential for scoring these tasks.
• Manual transcription is time-consuming and labor-intensive.

Purpose of the Study:

• To compare verbal fluency scores from manual transcriptions with those from automatic speech recognition (ASR) enhanced by machine learning.
• To evaluate the feasibility of using machine learning to improve ASR for verbal fluency scoring.

Main Methods:

• 1400 individuals' verbal fluency recordings (animal and letter F tasks) were automatically transcribed using Amazon Web Services.
• Automatic transcriptions were manually corrected to create "gold standard" transcriptions.
• Machine learning classifiers were trained to distinguish valid from invalid utterances for automated scoring.

Main Results:

• Machine learning classifiers achieved good separation of valid and invalid utterances for both tasks.
• Verbal fluency scores derived from automatic transcriptions showed high correlation with those from manual corrections.
• Some scores exhibited low correlations, indicating areas for ASR improvement.

Conclusions:

• Machine learning can enhance off-the-shelf ASR for scoring verbal fluency word lists.
• Automatically derived verbal fluency scores may be suitable for certain applications.
• Further advancements in ASR are necessary for reliable, fully automated verbal fluency assessment.