Investigation of the robustness of a non-uniform filterbank for cognitive load classification

Most of the current automatic speech-based cognitive load measurement systems utilize acoustic features estimated using a mel filterbank. However, a previous study showed that a non-uniform filterbank designed specifically to emphasize cognitive load information present in low frequencies was more effective than a mel filterbank under noise-free conditions. This paper investigates the effectiveness of the non-uniform filterbank and compares it with mel, Bark and Equivalent Rectangular Bandwidth (ERB) filterbank, under both clean and noisy conditions. Experiments are carried out both on the Stroop test and the Reading and Comprehension databases. Results show that the proposed filterbank consistently outperforms the other ones under both clean and noisy conditions. Specifically, in experiments performed on the Reading and Comprehension database the use of the proposed filter bank resulted in a relative reduction of error rates of 18.6% and 9.8% when compared to the commonly used mel filterbank under clean and noisy conditions respectively.