The Influence of Prosodic-Acoustic Features in Automatic Speech Recognition of Two Inland Brazilian Dialects: A Pilot Study

Abstract

This study investigates the role of prosodic-acoustic features in distinguishing two inland varieties of Brazilian Portuguese, Paraíba (PB) and São Paulo (SP), and examines how these features influence automatic speech recognition (ASR) performance. Grounded in sociophonetic theory and dialect-aware speech modeling, this pilot study evaluates whether prosodic-acoustic descriptors of duration, fundamental frequency (f0), and intensity provide robust cues for dialect differentiation and to what extent automatic dialect classification can be explained by the relative contribution of these acoustic dimensions. The study integrates two complementary methodological stages. Methodology 1 encompasses corpus design, acoustic extraction procedures, and inferential statistical modeling. Twenty university students (ten per dialect) produced controlled reading recordings, yielding 460 speech samples. Prosodic-acoustic parameters were extracted and analyzed using Linear Mixed-Effects Models to identify statistically significant dialect differences while controlling for inter-speaker variability. Methodology 2 applies automatic dialect classification using six machine-learning classifiers to evaluate the discriminative capacity of those features. The results are presented in two stages. Results 1 demonstrate that duration-related measures—particularly pause variability—provide the strongest explanatory power for distinguishing PB and SP, followed by intensity-related parameters such as spectral tilt and intensity variation coefficient, whereas f0 measures show comparatively smaller effects. Results 2 show that classification accuracy ranges from 94.6% to 97.8%, with tree-based models achieving the highest performance. Feature-importance analyses reveal a hierarchical contribution of acoustic parameters dominated by rhythmic-temporal organization. The convergence between explanatory statistical modeling and predictive classification demonstrates that inland dialect differentiation in Brazilian Portuguese is acoustically structured, prosodically organized, and computationally detectable. These findings contribute to the sociophonetic description of inland Brazilian Portuguese varieties and provide empirical grounding for incorporating dialect-sensitive prosodic modeling into ASR systems.