Exploring Pathological Speech Quality Assessment with ASR-Powered Wav2Vec2 in Data-Scarce Context

TL;DR

The paper addresses automatic speech quality assessment in data-scarce, pathological-speech contexts by adopting an audio-level regression approach using Wav2Vec2 features. It systematically compares self-supervised (SSL) and ASR-pretrained Wav2Vec2 models, finding that ASR-pretrained features (particularly 3K-ASR) yield the best intelligibility ($MSE=0.73$) and severity ($MSE=1.15$) predictions with only 95 training samples. Cross-domain evaluation on AHN Parkinsonian speech demonstrates strong generalization, with $MSE_{int}=0.22$ and $MSE_{sev}=0.37$, suggesting a meaningful link between ASR representations and speech quality assessment. The study also analyzes the impact of segment content and duration, showing that longer content improves performance while content differences between readings have limited effect on final decisions, highlighting the practical potential for data-efficient, audio-level evaluation in clinical contexts.

Abstract

Automatic speech quality assessment has raised more attention as an alternative or support to traditional perceptual clinical evaluation. However, most research so far only gains good results on simple tasks such as binary classification, largely due to data scarcity. To deal with this challenge, current works tend to segment patients' audio files into many samples to augment the datasets. Nevertheless, this approach has limitations, as it indirectly relates overall audio scores to individual segments. This paper introduces a novel approach where the system learns at the audio level instead of segments despite data scarcity. This paper proposes to use the pre-trained Wav2Vec2 architecture for both SSL, and ASR as feature extractor in speech assessment. Carried out on the HNC dataset, our ASR-driven approach established a new baseline compared with other approaches, obtaining average $MSE=0.73$ and $MSE=1.15$ for the prediction of intelligibility and severity scores respectively, using only 95 training samples. It shows that the ASR based Wav2Vec2 model brings the best results and may indicate a strong correlation between ASR and speech quality assessment. We also measure its ability on variable segment durations and speech content, exploring factors influencing its decision.

Figures (5)

• Figure 1: Scatter plot of intelligibility prediction
• Figure 2: Scatter plot of severity prediction
• Figure 3: Train and validation loss (MSE) curves from a random fold
• Figure 4: Model behavior at the segment level
• Figure 5: Absolute error variation across different segment durations