Modeling prosody in speech processing Free

Prosody is an important aspect of speech. In spoken language processing, effective prosody modeling helps to identify additional information beyond words (i.e., ‘‘how it is said’’ instead of ‘‘what is said’’) and thus better understand speech. In this talk, we will discuss how prosodic information is utilized in various speech processing tasks. Prosodic features are extracted to represent duration, pitch, and energy, with different normalization, and modeled using machine learning techniques. Research has shown that prosody provides valuable information for tasks such as automatic identification of important events in spoken language (e.g., sentence boundaries or punctuation, disfluencies, discourse markers, topics, and emotions in dialog). These phenomena are important for enriching speech recognition output and helping downstream language processing modules. Modeling prosodic variation across speakers is also useful for these tasks, as well as for developing speaker recognition systems. Additionally, some issues in machine learning techniques in prosody modeling will be discussed. Understanding how prosody is used to signal interesting events in speech will help to build better synthesis models for generating more natural and expressive speech.