Talking Soundscapes: Automatizing Voice Transformations for Crowd Simulation

The addition of a crowd in a virtual environment, such as a game world, can make the environment more realistic. While researchers focused on the visual modeling and simulation of a crowd, its sound production has received less attention. We propose the generation of the sound of a crowd by retrieving a very small set of speech snippets from a user-contributed database, and transforming and layering voice recordings according to the character localization in the crowd simulation. Our proof-of-concept integrates state-of-the-art audio processing and crowd simulation algorithms. The novelty resides in exploring how we can create a flexible crowd sound from a reduced number of samples, whose acoustic characteristics (such as people density and dialogue activity) could be modeled in practice by means of pitch, timbre and time-scaling transformations.