Computational Methods for Assisting Radio Drama Production.

Abstract

Radio Drama is a theatrical form of art that usually exists solely in the acoustic domain consisting of music, speech, and sound effects and is most often consumed through broadcast radio. This thesis proposes methods for assisting a human creator in producing radio dramas. Much research has been done to aiding creativity using artificial intelligence techniques in storytelling, music composition, the visual arts, and fi lm. Despite that, radio drama is under-represented in such research. Radio drama consists of both literary aspects, such as plot, story characters, or environments, as well as production aspects, such as speech, music, and sound effects. While plenty of research has been examining each of those aspects individually there is currently no research that combines such studies in the context of radio drama production. In this thesis, an interdisciplinary approach to assisting a human creator in radio drama production is developed. The task is explored through the joint prism of natural language processing, music information retrieval, and automatic mixing. We show that individual literary aspects of radio drama can be automatically extracted from a story draft provided by a human creator, by using natural language processing methods. Formal rules can be used to express the aforementioned elements in the form of a script able to be read and altered by both the human creator and the computer. We devise recommender systems for sound, music, and audio effects to retrieve the assets required for production. Rules derived from radio drama literature can then use those recorded assets to produce a radio drama mix in a semi-automatic way. Furthermore, an adaptive reverberation effect suggests reverberation settings for each track based on track content and past user choices. The degree of success for individual tasks in aiding production is demonstrated using examples of radio drama production from raw stories and validated through objective evaluation metrics, and listening tests.