The recording environment has a significant impact on implementation and results. Environmental factors affecting the recorded signals include: noise level, saturation, and timing. As the noise level increases, it becomes more difficult for the AR model to distinguish the formant peaks from the noise floor. Time shifts (i.e. where do the signals start?) and signal duration both had decided impacts on all the methods investigated. The time issues are non-trivial. Ad hoc methods to account for timing variations from signal to signal had varying results. Dynamic Time Warping is one technique to address timing misalignment, but was not included in the scope of this project.
While the Power method is robust, it has little room for refinement. It's simplicity is, in fact, its limiting factor. On the other hand, small modifications to the Formant method do yield improved results. For the results obtained, however, this method is non-trivial to program. The Cepstrum method outperformed both of the other methods by more than 10x improvement in error rate. While the theory behind cepstral analysis is more advanced, the implementation is comparable to the Formant method.