Speech production is basically
a source-filter model. The source is the air provided by the lungs.
The filter is the spectral shaping performed by the vocal tract. The
convolution of the two in the time domain produces the desired utterance.
Because these are two separate processes, the source excitation and
the filter implementation can be analyzed separately. Optimizing both
the source sub-model and the filter sub-model can improve our speech
utterance.
Source Excitation
The source excitation can
be one of two types. For voiced speech, the vocal folds close and open
rhythmically to make the air from the lung into an "impulse train."
This impulse creates the pitch for a sustained voiced sound. For unvoiced
speech, the source is white noise. In this case, the air that flows
out of the lungs and between the tongue and mouth produces a relatively
random sound.
Acoustic Tube and Transmission
Line Model
In its simplest form, vocal
tract can be modeled as a lossless acoustic tube. The cross-sectional
area of the tube and the speed of the air determine the sound pressure
and volume velocity, which in turn determines the output speech. The
vocal tract can also be modeled as a transmission line. The acoustical
resistance, mass, and compliance are distributed along the tube in the
same manner as the resistance, inductance, and capacitance along a transmission
line.

The single acoustic tube/transmission
line model is not adequate to model the wide range of sounds we create.
Since speech production is characterized by changing vocal-tract shape,
it is more appropriate to create multiple acoustic tubes or cascading
transmission lines as models. The specific shape of a vocal-tract and
how it changes in time determines the actual word utterances we perceive
as speech.

The following figure shows
how the source/filter approach applies to the human vocal production
organs. The lungs, vocal folds and trachea all belong to the "source"
side of the model. The various cavities, the velum and the tongue hump
are part of the filter end.
