next up previous
Next: Minimisation Up: Simulations Previous: Simulations


Speech data was used for the simulations: 30 s of Finnish speech was digitised with 16 kHz sampling rate and high-pass filtered. Power spectra were computed every 8 ms using short time Fourier transformations with Hamming windows of length 16 ms. This results in 3749 vectors of dimension 128. Energy was computed for 34 channels whose spacing imitates the frequency scale of human ear. Logarithms were taken from the energies after adding small constants. The final data thus consisted of 3749 vectors of dimension 34.

This particular preprocessing was chosen because it is typical in current speech recognition systems.

Harri Lappalainen