Speech segmentation with a neural encoder model of working memory

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1070–1080 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Speech segmentation with a neural encoder model of working memory Micha Elsner and Cory Shain melsner0@gmail.com and shain.3@osu.edu Department of Linguistics The Ohio State University, Columbus, OH 43210 Abstract We present the first unsupervised LSTM speech segmenter as a cognitive model of the acquisition of words from unseg- mented input. Cognitive biases toward phonological and syntactic predictability in speech are rooted in the limitations of human memory (Baddeley et al., 1998); compressed representations are easier to acquire and retain in memory. To model the biases introduced by these memory limitations, our system uses an LSTM- based encoder-decoder with a small num- ber of hidden units, then searches for a segmentation that minimizes autoencod- ing loss. Linguistically meaningful seg- ments (e.g. words) should share regu- lar patterns of features that facilitate de- coder performance in comparison to ran- dom segmentations, and we show that our learner discovers these patterns when trained on either phoneme sequences or raw acoustics. To our knowledge, ours is the first fully unsupervised system to be able to segment both symbolic and acous- tic representations of speech. 1 Introduction This paper describes a new cognitive model of the acquisition of word-like units from unsegmented input. The model is intended to describe the pro- cess by which pre-linguistic infants learn their ear- liest words, a stage they pass through during the first year of life (Jusczyk and Aslin, 1995; Bergel- son and Swingley, 2012). Our model is based on the standard memory model of Baddeley and Hitch (1974) in which the listener encodes lexical items into phonological working memory, but rep- resents the entire sentence as a higher-level syn- tactic structure without phonological detail. Our model implements this architecture using encoder- decoder LSTMs with limited memory capacity, then searches for word segmentations which make it easy to remember the sentence.1 Word learning has been extensively studied in previous research, both with transcribed symbolic input and acoustics. Why attempt yet another ap- proach? Our model has three main advantages. First, as a cognitive model, it relates the kinds of learning biases used in previous work to the wider literature on working memory. Second, its sequence-to-sequence neural architecture allows it to handle either one-hot symbolic input or dense vectors of acoustic features. In contrast, existing models are typically designed for “clean” sym- bolic input, then retrofitted with additional mech- anisms to cope with acoustics. Finally, neural net- works have been impressively successful in super- vised language processing domains, yet are still underused in unsupervised learning. Even sys- tems which do use neural nets to model lexical ac- quisition generally require an auxiliary model for clustering the embeddings, which can make their learning objectives difficult to understand. Our system uses the well-understood autoencoder ob- jective to perform the segmentation task without requiring auxiliary clustering, and thus suggests a new direction for neural unsupervised learning. In an experiment …

Content truncated to 500 words