Voice adaptation by color-encoded frame matching as a multi-objective optimization problem for future games

Abstract

Voice adaptation is an interactive speech processing technique that allows the speaker to transmit with a chosen target voice. We propose a novel method that is intended for dynamic scenarios, such as online video games, where the source speaker’s and target speaker’s data are nonaligned. This would yield massive improvements to immersion and experience by fully becoming a character, and address privacy concerns to protect against harassment by disguising the voice. With unaligned data, traditional methods, e.g., probabilistic models become inaccurate, while recent methods such as deep neural networks (DNN) require too substantial preparation work. Common methods require multiple subjects to be trained in parallel, which constraints practicality in productive environments. Our proposal trains a subject nonparallel into a voice profile used against any unknown source speaker. Prosodic data such as pitch, power and temporal structure are encoded into RGBA-colored frames used in a multi-objective optimization problem to adjust interrelated features based on color likeness. Finally, frames are smoothed and adjusted before output. The method was evaluated using Mean Opinion Score, ABX, MUSHRA, Single Ease Questions and performance benchmarks using two voice profiles of varying sizes and lastly discussion regarding game implementation. Results show improved adaptation quality, especially in a larger voice profile, and audience is positive about using such technology in future games.