We found that although the
LPC model is greatly simplified, it is still useful in allowing us to
efficiently describe vocal tract behavior. We used an 18 order LPC filter
model, which was a compromise between quality and complexity. More poles
would yield better results but only to a certain point. From
there, further improvements
to the other parts
of the model are needed.
Our samples did not come
out very smooth. This is due to our simple concatenation method, making
the separate phonemes sound choppy and disconnected.
The square pulse train that
we generated to use as a source turned out to be lacking in similarity
to a vocal cord excitation. The sound coming from our vocal cords is
more gradual and less sharp-cornered than the square pulse train. The
result is speech that sounded unnatural, with a sharp mechanical quality.
Even though we did make our
computer say a few words, it would not be mistaken for a human voice.