Normalization through Fine-tuning: Understanding Wav2vec 2.0 Embeddings for Phonetic Analysis

Phonetic normalization is essential in speech recognition and analysis, ensuring the comparability of features extracted from raw audio data. In the current paradigm of fine-tuning pretrained large transformer models, however, it is not treated as a necessary preprocessing step; rather, it is implicitly carried out within the models themselves. This study examines how normalization emerges in transformer-based speech models, with a focus on Wav2vec2.0. Understanding this process is a crucial step toward effectively leveraging model embeddings as phonetic features for speech analysis. Our results show that fine-tuning Wav2vec2.0 can achieve phonetic normalization by selectively suppressing irrelevant information, such as speaker sex, while enhancing task-relevant information, such as tones and finals. In multi-task learning, however, the embeddings of fine-tuned models retain information for each task without compromising performance, suggesting that suppressing task-irrelevant information is not strictly required for effective classification. These results demonstrate that phonetic normalization can be flexibly achieved in speech models, providing insights into potential parallels in human speech perception.