0.4 5 - voice recognition: graphs

1. This time-domain signal shows clear segmentation between different numbers.

2．This shows a normalized spectrogram of the number “1”. Formants are visible but not clear.

3. Enhanced by a non-linear filter to emphasize the difference between peak values and background. Formants are much more distinct.

4. This shows the filtered spectrum in the Mel-scale, a logarithmic scale that models human hearing. Corner-frequency of 700 Hz used.

5. Weighted scatter plot with contours of the maximum likelihood GMM overlaid, showing the formants.

6. The GMM generated by the Mel-scale signal differs greatly from the linear-frequency version.