TABLE OF CONTENTS The source-filter model is perhaps the ultimate in speech parameterisation, with different processing blocks dedicated to replicating the effects of the human vocal system: LPC/LSP for the vocal tract, random noise (and similar) for the lung excitation, and a pitch filter or similar to recreate the effect of the glottis.
Measurements of the human pitch-production system, especially those using microwave and X-ray sensors, reveal the action of the glottis, which is not a smooth action: it does not generate a pure sinewave tone. In actual fact, the pitch waveform is made up of a sequence of very spiky pulses. This is shown in Figure 5.14 where one pulse has been identified from a sequence of several plotted as if isolated from a speech utterance.
There has been quite a lot of research on determining pitch shapes: how these relate to overall vocal quality, speech intelligibility, and so on. There is substantial evidence that the fidelity of the pitch pulse shape is important to overall perceived quality, and other evidence to indicate that the specific pulse shapes, which vary considerably from person to person, are one of the differentiating factors for speaker recognition (where an automatic system identifies someone through their voice, see Section 7.5).
When coding or compressing speech in a parametric fashion, there are several items of information that are important for...
