Public defence in Acoustics and Speech Technology, M.Sc. Manila Kodali

Title of the thesis: Speech-based classification and regression studies on vocal intensity and severity level of Parkinson’s disease

Thesis defender: Manila Kodali
Opponent: Prof. John Hansen, University of Texas at Dallas, US
Custos: Prof. Paavo Alku, Aalto University School of Electrical Engineering 

Speech is a fundamental form of human communication conveying linguistic, paralinguistic,
and extralinguistic cues. Advances in machine learning (ML) have enabled automatic analysis of
such information, offering a scalable and low-cost alternative to traditional clinical voice
assessment. This thesis focuses on two interlinked topics, vocal intensity and Parkinson’s
disease (PD). The thesis studies the multi-class classification of vocal intensity categories, the
prediction of sound pressure level (SPL) from amplitude-normalized speech, and the multi-
class classification of PD severity using spectral features, pre-trained embeddings, and wavelet
scattering representations with ML models.

For the vocal intensity category classification problem, the best system improved accuracy from the chance level (of 25 %) up to almost 86 %. For the SPL prediction problem, the best regression system was able to estimate the ground truth SPL values from speech signals produced by pathological and healthy speakers with a mean absolute error of approximately 2.0 dB. In the severity classification of PD, monologue and reading tasks outperformed vowel and sentence production tasks, as well as a diadochokinetic task. Articulation and fusion feature sets performed significantly better than phonation and prosody features, while the choice of classifier had only a minor impact on accuracy.

In conclusion, spectral features, and pre-trained model embeddings, combined with ML models, showed improved performance compared to non-fine tuned embeddings for both of the studied vocal intensity problems. In particular, the prediction of SPL from amplitude-normalized speech signals stands out as an attractive new method with strong clinical potential because it enables improved interpretability of speech-based biomarking technology when speech is recorded in non-calibrated real-world conditions. Furthermore, the thesis shows that a monologue speaking task combined with articulation and fusion features yielded the highest accuracy in the classification of the severity level of PD.

Key words: Vocal intensity, Parkinson's disease, sound pressure level, machine learning, classification, regression

Thesis available for public display 7 days prior to the defence at Aalto University's public display page.

Contact: 
Phone: +358505713343
www.linkedin.com/in/manila-kodali-7bb346150