Automatic Evaluation of the Pataka Test Using Machine Learning and Audio Signal Processing

This study presents an automated methodology for evaluating the Pataka test, a key assessment of syllable repetition speed in motor speech disorders. The proposed system employs the Wav2Vec2 model for feature extraction and transcription, supported by audio preprocessing steps (resampling, mono conversion, normalization) that enhance data consistency. Sixteen audio recordings, each containing repeated syllables from two participants, were analyzed by the automated tool and by three speech pathologists. Ground truth measurements were established using mel spectrograms for syllable boundaries, yielding reliable benchmarks of repetition rates. Results revealed that the machine learning model significantly outperformed manual raters: the mean squared error (MSE) was 0.07 for the model versus 1.18 for human evaluations. These findings underscore the potential clinical benefits of automated tools in improving diagnostic precision and reducing variability in speech evaluations. Future work should address the system’s tendency to miss occasional syllables and extend the dataset to include a wider demographic range. Overall, this study demonstrates that integrating deep learning with audio signal processing offers a scalable, objective alternative to traditional Pataka test assessments.