Using Spectral Analysis to Match Players Voice to in Game Characters
By: Nimrod Barak
Supervisor: Gary Keith Brubaker
Masters of Interactive Technology degree conferred June 12, 2009
Thesis / Project completed: June 10, 2009

The goal of this project was to test whether significant voice transformations that the majority of users do not detect could be done in real-time. It also studied the nature and tradeoffs of voice transformations. The thesis employed Fast Fourier Transformation and spectral analysis techniques and combined them with current theories on the human voice. Although, voice transformation technology currently exists, users report that they can tell when real-time applications, such as XBOX Live voice masking, change a voice. Thus, it is important to test and explore how this technology could be improved.

To test the hypothesis, this project uses a spectral analyzer program that processes sound data and displays its frequency information in spectrograph form. It then manipulates the information to apply low-noise transforms quickly. The program can apply low-pass and high-pass filters, sharpen, clarify, attenuate, shift pitch, and shape inflection inside frequency ranges or to the whole sound bite.

A series of sound tests were conducted to confirm the naturalness of the sounds post-transformations. The generation speed was empirically tested as well as analyzed for Big O limiting behavior, which is the rate that computational time increases based on the amount of data that is processed. Testing showed that transformations which reallocated energy went undetected by the majority of listeners, and that transformations which moved frequency data over time created noise artifacts that require more computational time to deal with. Since this program was able to apply theses transformations quickly without major optimization, this thesis will show that real-time, unnoticed transforms are valid and in fact inevitable.

Download entire thesis (.pdf)