Since the experiments of Saffran et al. (1996a), there has been a great deal of interest in the question of how statistical regularities in the speech stream might be used by infants to begin to identify individual words. In this work, we use computational modeling to explore the effects of different assumptions the learner might make regarding the nature of words -in particular, how these assumptions affect the kinds of words that are segmented from a corpus of transcribed child-directed speech. We develop several models within a Bayesian ideal observer framework, and use them to examine the consequences of assuming either that words are independent units, or units that help to predict other units. We show through empirical and theoretical results that the assumption of independence causes the learner to undersegment the corpus, with many two-and three-word sequences (e.g. what's that, do you, in the house) misidentified as individual words. In contrast, when the learner assumes that words are predictive, the resulting segmentation is far more accurate. These results indicate that taking context into account is important for a statistical word segmentation strategy to be successful, and raise the possibility that even young infants may be able to exploit more subtle statistical patterns than have usually been considered.

What mechanisms support the ability of human infants, adults, and other primates to identify words from fluent speech using distributional regularities? In order to better characterize this ability, we collected data from adults in an artificial language segmentation task similar to in which the length of sentences was systematically varied between groups of participants. We then compared the fit of a variety of computational modelsincluding simple statistical models of transitional probability and mutual information, a clustering model based on mutual information by Swingley (2005), PARSER (Perruchet & Vintner, 1998), and a Bayesian model. We found that while all models were able to successfully complete the task, fit to the human data varied considerably, with the Bayesian model achieving the highest correlation with our results.

Infants segment words from fluent speech during the same period when they are learning phonetic categories, yet accounts of phonetic category acquisition typically ignore information about the words in which sounds appear. We use a Bayesian model to illustrate how feedback from segmented words might constrain phonetic category learning by providing information about which sounds occur together in words. Simulations demonstrate that word-level information can successfully disambiguate overlapping English vowel categories. Learning patterns in the model are shown to parallel human behavior from artificial language learning tasks. These findings point to a central role for the developing lexicon in phonetic category acquisition and provide a framework for incorporating top-down constraints into models of category learning.

Before they even speak, infants become attuned to the sounds of the language(s) they hear, processing native phonetic contrasts more easily than nonnative ones. For example, between 6 to 8 mo and 10 to 12 mo, infants learning American English get better at distinguishing English and [l], as in “rock” vs. “lock,” relative to infants learning Japanese. Influential accounts of this early phonetic learning phenomenon initially proposed that infants group sounds into native vowel- and consonant-like phonetic categories—like and [l] in English—through a statistical clustering mechanism dubbed “distributional learning.” The feasibility of this mechanism for learning phonetic categories has been challenged, however. Here, we demonstrate that a distributional learning algorithm operating on naturalistic speech can predict early phonetic learning, as observed in Japanese and American English infants, suggesting that infants might learn through distributional learning after all. We further show, however, that, contrary to the original distributional learning proposal, our model learns units too brief and too fine-grained acoustically to correspond to phonetic categories. This challenges the influential idea that what infants learn are phonetic categories. More broadly, our work introduces a mechanism-driven approach to the study of early phonetic learning, together with a quantitative modeling framework that can handle realistic input. This allows accounts of early phonetic learning to be linked to concrete, systematic predictions regarding infants’ attunement.

We present a simple approach to improve direct speech-to-text translation (ST) when the source language is low-resource: we pre-train the model on a high-resource automatic speech recognition (ASR) task, and then fine-tune its parameters for ST. We demonstrate that our approach is effective by pre-training on 300 hours of English ASR data to improve Spanish-English ST from 10.8 to 20.2 BLEU when only 20 hours of Spanish-English ST training data are available. Through an ablation study, we find that the pre-trained encoder (acoustic model) accounts for most of the improvement, despite the fact that the shared language in these tasks is the target language text, not the source language audio. Applying this insight, we show that pre-training on ASR helps ST even when the ASR language differs from both source and target ST languages: pre-training on French ASR also improves Spanish-English ST. Finally, we show that the approach improves performance on a true low-resource task: pre-training on a combination of English ASR and French ASR improves Mboshi-French ST, where only 4 hours of data are available, from 3.5 to 7.1 BLEU.

Deep neural networks (DNNs) have become a standard component in supervised ASR, used in both data-driven feature extraction and acoustic modelling. Supervision is typically obtained from a forced alignment that provides phone class targets, requiring transcriptions and pronunciations. We propose a novel unsupervised DNN-based feature extractor that can be trained without these resources in zeroresource settings. Using unsupervised term discovery, we find pairs of isolated word examples of the same unknown type; these provide weak top-down supervision. For each pair, dynamic programming is used to align the feature frames of the two words. Matching frames are presented as input-output pairs to a deep autoencoder (AE) neural network. Using this AE as feature extractor in a word discrimination task, we achieve 64% relative improvement over a previous stateof-the-art system, 57% improvement relative to a bottom-up trained deep AE, and come to within 23% of a supervised system.

Developing better methods for segmenting continuous text into words is important for improving the processing of Asian languages, and may shed light on how humans learn to segment speech. We propose two new Bayesian word segmentation methods that assume unigram and bigram models of word dependencies respectively. The bigram model greatly outperforms the unigram model (and previous probabilistic models), demonstrating the importance of such dependencies for word segmentation. We also show that previous probabilistic models rely crucially on suboptimal search procedures.

Zero-resource speech technology is a growing research area that aims to develop methods for speech processing in the absence of transcriptions, lexicons, or language modelling text. Early term discovery systems focused on identifying isolated recurring patterns in a corpus, while more recent full-coverage systems attempt to completely segment and cluster the audio into word-like units-effectively performing unsupervised speech recognition. This article presents the first attempt we are aware of to apply such a system to large-vocabulary multi-speaker data. Our system uses a Bayesian modelling framework with segmental word representations: each word segment is represented as a fixed-dimensional acoustic embedding obtained by mapping the sequence of feature frames to a single embedding vector. We compare our system on English and Xitsonga datasets to state-of-the-art baselines, using a variety of measures including word error rate (obtained by mapping the unsupervised output to ground truth transcriptions). Very high word error rates are reported-in the order of 70-80% for speaker-dependent and 80-95% for speaker-independent systems-highlighting the difficulty of this task. Nevertheless, in terms of cluster quality and word segmentation metrics, we show that by imposing a consistent top-down segmentation while also using bottomup knowledge from detected syllable boundaries, both singlespeaker and multi-speaker versions of our system outperform a purely bottom-up single-speaker syllable-based approach. We also show that the discovered clusters can be made less speakerand gender-specific by using an unsupervised autoencoder-like feature extractor to learn better frame-level features (prior to embedding). Our system's discovered clusters are still less pure than those of unsupervised term discovery systems, but provide far greater coverage.