CAREER: Modeling Spoken Language Without Parallel Text Annotations

点击翻译按钮获取中文摘要

Automatic speech recognition and understanding technology has been widely adopted into personal digital assistants, automatic transcription of videos and meetings, and many more applications. Building these systems requires massive datasets of speech audio that is human-transcribed into text. When sufficient data is available for a particular domain, modern models based on deep neural networks are capable of highly accurate speech recognition and downstream language understanding tasks. However, for the vast majority of the world's 7,000 languages and even more numerous dialects, large scale annotated datasets simply do not exist, preventing speech technology from serving these languages and their speakers. Inspired by the fact that humans learn to speak long before they can read or write, this CAREER project explores a new paradigm for speech processing that does not rely on transcribed speech. Instead, it develops new models that are capable of learning spoken language directly from speech audio, and applies these models to tasks including building speech recognizers without transcribed speech and automatically translating speech from one language into another. These advances fit within a larger movement in the research community to dramatically reduce the cost and increase the availability of speech recognition and understanding technology to many more languages and users than are served today.This project leverages self-supervised and multimodal learning approaches to automatically discover linguistic structure (phones, words, phrases, etc.) in the raw speech signal which can be treated as ``pseudo-text'' and used in place of conventional text for downstream tasks. It develops new neural network layers for attention-based segmentation of speech, applied in a hierarchical fashion to discover speech units at multiple levels of abstraction. A second novel technique involves adding self-prediction layers and training objectives to a model using the segmentation layers, where the higher layers that would capture word-like structure attempt to predict the tokenization of lower layers that capture sub-word structure. In this way, the model can automatically learn a pronunciation lexicon that captures the compositional relationship between the different tiers of discovered speech units. The project applies these techniques to three downstream applications that are steadily growing in importance in the speech field: unsupervised speech recognition, textless speech-to-speech translation, and textless generation speech for dialog and image captioning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.