Bottom-up learning of a phonetic system using an autoencoder

Abstract

<p>This study investigates an autoencoder model’s effectiveness in learning phonemes and distinctive features from unsegmented, non-transcribed wave data, mirroring infants’ early language acquisition stages. The results of the experiment on Mandarin and English show that features can be learned through repeated projection of sounds to a hidden space and reconstruction back to their original form without prior knowledge of segmentation. The acquired knowledge is represented by different distributions in the hidden space. The model was capable of clustering segments of the same phoneme and projecting different phonemes to separate regions in the hidden space. However, the model struggled to cluster allophones closely together, which may indicate the boundary between the contributions of bottom-up and top-down information. Overall, the study suggests that sound knowledge can be learned to a certain degree through unsupervised learning techniques that do not require labeled data or prior knowledge of segmentation, and that this approach may provide insights into the early stages of human language acquisition.</p>