Unsupervised neural and Bayesian models for zero-resource speech processing
In settings where only unlabelled speech data is available, zero-resource speech technology needs to be developed without transcriptions, pronunciation dictionaries, or language modelling text. There are two central problems in zero-resource speech processing: (i) finding frame-level feature representations which make it easier to discriminate between linguistic units (phones or words), and (ii) segmenting and clustering unlabelled speech into meaningful units. In this thesis, we argue that a combination of top-down and bottom-up modelling is advantageous in tackling these two problems. To address the problem of frame-level representation learning, we present the correspondence autoencoder (cAE), a neural network trained with weak top-down supervision from an unsupervised term discovery system. By combining this top-down supervision with unsupervised bottom-up initialization, the cAE yields much more discriminative features than previous approaches. We then present our unsupervised segmental Bayesian model that segments and clusters unlabelled speech into hypothesized words. By imposing a consistent top-down segmentation while also using bottom-up knowledge from detected syllable boundaries, our system outperforms several others on multi-speaker conversational English and Xitsonga speech data. Finally, we show that the clusters discovered by the segmental Bayesian model can be made less speaker- and gender-specific by using features from the cAE instead of traditional acoustic features. In summary, the different models and systems presented in this thesis show that both top-down and bottom-up modelling can improve representation learning, segmentation and clustering of unlabelled speech data.