Linear Prediction Analysis of Speech
Step1: Select one of the audio/waveform file provided speech utterances from the given dropdown.
Step2: Click on the 'Generate Spectrum' button. The 'Windowed Waveform' signal, the 'Log Spectrum', the 'LP Residual' and the 'LP Spectrum' are displayed.
Step3: Play the short segment (20-50 ms) of voiced speech (say a vowel), using 'Play/Pause' button, and compute the short-time.
Step4: Increase the 'LP Order' from the default value 1, to values provided in the drop down box. Observe the 'LP residual' and 'LP log spectrum'.
Step5: Observe that the 'LP log spectrum' approximates the short-time spectral envelope (i.e., the formants) better with increased 'LP order'.
Step6: As the 'LP order' is increased to a very large value, say 40, the 'LP spectrum' tries to approximate the short-time spectral envelope more closely thereby approximating peaks due to pitch harmonics.
Step7: Study the effect of pre emphasis flag on the 'LP spectrum'.
Step8: Observe the compute the 'LP residual' signal for different 'LP orders'
Step9: Observe the correlations in the input segment of speech. Notice it being gradually removed in the 'LP residual' signal as the 'LP order' is increased.
Step10: Observe that the 'LP residual' signal for an 'LP order' of 10 or above removes most of the correlations in the input signal, and the resulting signal looks like a train of impulses.
Step11: Repeat the experiment by selecting an unvoiced segment of speech, say a fricative /s/.
Step12: Write a brief note on the observations.