Speech-Based Dysarthria Severity Classification in Amyotrophic Lateral Sclerosis: A Controlled Deep Learning Benchmark

This study examines whether adding waveform-derived information-theoretic descriptors to Mel-spectrogram inputs offers a measurable benefit for five-class amyotrophic lateral sclerosis (ALS) dysarthria speech-status classification. The primary evidence is a study-specific, five-fold recording-level CNN comparison: Mel-only accuracy is 0.926±0.012, while historical fusion accuracy is 0.927±0.008. A distinct CNN fusion verification/reproduction run, with traced propagation of the Mel input and all 17 information-theoretic descriptors, reports accuracy of 0.927±0.010; it does not retrospectively verify the archived historical input tensors. Thus, the available comparisons show only a small numerical change, not a demonstrated general advantage of fusion. Complementary analyses include a direct participant-disjoint CNN comparison, a fixed-duration CNN comparison, and seed, duration, padding, and task-specific controls. In the participant-disjoint comparison, Mel-only and fusion achieve accuracies of 0.820 and 0.823, respectively, whereas under the fixed-duration condition they achieve 0.914 and 0.915. These distinct result series show similarly small numerical differences between Mel-only and fusion under the additional evaluation conditions. They do not establish a matched estimate of the effect of participant overlap or a statistically reliable advantage of feature fusion. Source-code auditing identifies discrepancies between an available implementation and the documented score-producing benchmark, without establishing that leakage occurred in the latter. Missing run-linked records, class imbalance, and the absence of external validation constrain conclusions to the evaluated speech-processing pipelines. Class 4 represents ALS without dysarthria, not early bulbar disease detection.

Authors

Institutions

Publication Details

Journal
Electronics
Published
2026-11-17
DOI
https://doi.org/10.3390/electronics15225362
Primary Topic
Voice and Speech Disorders
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Speech-Based Dysarthria Severity Classification in Amyotrophic Lateral Sclerosis: A Controlled Deep Learning Benchmark

Stefano Caputo, Giacomo Borghini, Attaphongse Taparugssanagorn, Lorenzo Mucchi
1 links
Electronics
Voice and Speech Disorders
article

Speech-Based Dysarthria Severity Classification in Amyotrophic Lateral Sclerosis: A Controlled Deep Learning Benchmark

Stefano Caputo, Giacomo Borghini, Attaphongse Taparugssanagorn, Lorenzo Mucchi
article en

Abstract

This study examines whether adding waveform-derived information-theoretic descriptors to Mel-spectrogram inputs offers a measurable benefit for five-class amyotrophic lateral sclerosis (ALS) dysarthria speech-status classification. The primary evidence is a study-specific, five-fold recording-level CNN comparison: Mel-only accuracy is 0.926±0.012, while historical fusion accuracy is 0.927±0.008. A distinct CNN fusion verification/reproduction run, with traced propagation of the Mel input and all 17 information-theoretic descriptors, reports accuracy of 0.927±0.010; it does not retrospectively verify the archived historical input tensors. Thus, the available comparisons show only a small numerical change, not a demonstrated general advantage of fusion. Complementary analyses include a direct participant-disjoint CNN comparison, a fixed-duration CNN comparison, and seed, duration, padding, and task-specific controls. In the participant-disjoint comparison, Mel-only and fusion achieve accuracies of 0.820 and 0.823, respectively, whereas under the fixed-duration condition they achieve 0.914 and 0.915. These distinct result series show similarly small numerical differences between Mel-only and fusion under the additional evaluation conditions. They do not establish a matched estimate of the effect of participant overlap or a statistically reliable advantage of feature fusion. Source-code auditing identifies discrepancies between an available implementation and the documented score-producing benchmark, without establishing that leakage occurred in the latter. Missing run-linked records, class imbalance, and the absence of external validation constrain conclusions to the evaluated speech-processing pipelines. Class 4 represents ALS without dysarthria, not early bulbar disease detection.

ElectronicsVol. 15(22)
Asian Institute of Technology (TH), University of Florence (IT)
Good health and well-being
Voice and Speech Disorders
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon