Transformer-based speech models have recently been shown to perform well under impoverished and ambiguous supervision, a setting motivated by poverty-of-the-stimulus arguments, yet their learning dynamics remain unclear. In this work, we study this question with wav2vec2.0 on Mandarin tone classification, using 70 scale–noise configurations, capacity interventions, and perturbation probes to examine the dynamics of abstraction-consistent generalization and exception memorization, yielding representative regimes of Robust Generalization, Abrupt Transition, and High-noise Collapse. We find a consistent abstraction–exception trajectory: when phonological recovery is possible, the model first exhibits phonologically abstraction-consistent behavior and only later fits corrupted labels as empirical exceptions. Capacity analyses show that this trajectory is not merely a large-model artifact: it persists under truncated Transformer encoders, while acoustic rank compression modulates but does not selectively separate abstraction from exception memorization. Finally, acoustic perturbations show different local robustness: abstraction-consistent predictions remain stable under stronger waveform perturbations, whereas fitted exceptions behave as fragile local overrides on top of the broader abstraction-consistent behavior.

Lexical tones play a crucial role in Sino-Tibetan languages. However, current phonetic fieldwork relies on manual effort, resulting in substantial time and financial costs. This is especially challenging for the numerous endangered languages that are rapidly disappearing, often compounded by limited funding. In this paper, we introduce pitch-based similarity representations for tone transcription, named Tone2Vec. Experiments on dialect clustering and variance show that Tone2Vec effectively captures fine-grained tone variation. Utilizing Tone2Vec, we develop the first automatic approach for tone transcription and clustering by presenting a novel representation transformation for transcriptions. Additionally, these algorithms are systematically integrated into an open-sourced and easy-to-use package, ToneLab, which facilitates automated fieldwork and cross-regional, cross-lexical analysis for tonal languages. Extensive experiments were conducted to demonstrate the effectiveness of our methods.