next up previous contents
Next: The Autocorrelation from the Up: Short-Term Fourier Analysis  Previous: Homomorphic filtering

Mel scaled analysis

All the above analysis techniques place equal emphasis on every part of the frequency scale from zero up to the maximum representable frequency. As discussed in section 3, it is often better to place the bins on a non-linear frequency scale. Figure 32 shows the result of warping 23 onto the mel scale.

Figure 32: Mel scaled log power spectrum

Note that these plots have only shown part of the spectrum - as illustrating in figure 13 the DFT is periodic. To transform a plot such as figure 32 to the cepstral domain involves either reflecting about zero frequency or using the Discrete Cosine Transform.

Speech Vision Robotics group/Tony Robinson