Layer-wise Probing of wav2vec 2.0 and Whisper for Consonant Cluster Reduction in African American English

A new study probes how two widely used speech models internally represent consonant cluster reduction in African American English, finding that the models encode the phenomenon as structured gradient variation rather than simple deletion [1]. Researchers conducted layer-wise probing of wav2vec2-base and Whisper-small to examine how they handle consonant cluster reduction (CCR) in African American English (AAE), a phonological process that contributes to automatic speech recognition disparities [1][2]. The study employed two tasks: segmental reduction detection and segmental restoration of underlying cluster identity [2]. Both models distinguished reduced and canonical forms with high accuracy [1][2]. The findings indicate that reduced segments retain cues to their underlying stops, suggesting the models encode CCR as structured gradient phonological variation rather than treating it as outright segmental deletion [1][2]. This challenges simpler assumptions about how speech models process dialectal phonological patterns. The research addresses a gap in the literature. Self-supervised and supervised speech models are increasingly used to investigate which linguistic information their internal representations encode and at what level of abstraction they encode it [2]. Consonant cluster reduction in AAE had remained underexplored in this context, despite being a widespread phonological process [2]. The work was conducted as speaker-independent probing, meaning the models were tested on speakers not seen during training, which strengthens the generalizability of the results [2]. The paper was submitted to arXiv on June 22, 2026, under the Computation and Language category [1]. While the primary paper does not provide direct comparisons to commercial ASR performance metrics, the structured encoding finding carries implications for understanding why speech recognition systems may struggle with AAE. Prior research has documented that CCR is a source of ASR disparity [2]. The new evidence that models internally preserve gradient information about underlying stops suggests potential pathways for improving recognition accuracy on reduced forms without requiring explicit phonological rule programming. The study contributes to a broader effort to audit what linguistic knowledge speech models acquire during pre-training. By demonstrating that modern speech models encode structured phonological representations of AAE CCR patterns, the work provides a foundation for future research on dialect-aware speech technology [1][2].