Representation learning for word segmentation

Around 40% of natural languages have no standardized writing system and are under-resourced, making text-based language processing tools impractical for their speakers. This thesis develops models to document these languages and provide technology to the communities that speak them. Building on prior work, I focus on two approaches: translation systems trained to map text in a high-resource language to phone sequences of an unwritten language, and unsupervised learning from speech alone, without additional supervision. My hypothesis is that exploring different input representations can broaden the applicability of the translation-based method and improve the performance of the unsupervised one. The thesis is organized into three parts. The first asks whether translation-based segmentation models can work directly with raw speech rather than phonetic transcriptions. I find that their performance is limited when speech is the input modality, which motivates a shift towards unsupervised methods. The second part explores techniques for producing fixed-dimensional vector representations of speech segments, a key intrinsic task in unsupervised word segmentation. I apply mean pooling to frame-level, pre-trained self-supervised speech representations to obtain acoustic word embeddings (AWEs). On a standard word discrimination task using English HuBERT representations, this approach comes close to the state of the art for English at the time of writing. Although trained only on English data, these representations also perform competitively on Xitsonga, Mandarin, and French, and continuing pre-training with 50 hours of target-language data improves results further — evidence of HuBERT's robustness across diverse linguistic settings. While mean pooling performs well, prior work suggests that learning the pooling operation can do better. Learned pooling is typically trained on pairs of word-like units, which unsupervised systems mine using k-nearest neighbor (KNN) search — a slow and computationally expensive process. I address this by using a multilingual phone recognizer (MPR) to mine phone n-gram pairs for training the pooling function, and show that this approach is significantly faster than KNN-based mining while also performing better. The final part investigates whether self-supervised representations can improve word boundary estimation in unsupervised word segmentation. I use mean-pooled HuBERT frame-level representations with segmental k-means, which at the time was the state of the art on Task 2 of the ZeroSpeech Challenge and originally relied on MFCCs. Self-supervised representations consistently outperform subsampled MFCCs on precision and F1. The results also point to a correlation between acoustic word discrimination and segmentation performance, suggesting that word discrimination is a useful proxy for evaluating unsupervised word segmentation. Taken together, this work shows that unsupervised word segmentation is a better option than translation-based systems for unwritten languages, and that self-supervised representation techniques improve both acoustic word discrimination and unsupervised segmentation.