[20] L'lasons de Aether ver Beinags #2

Since I want the characters in my game to have some level of recognizably in their voice, I should try and not just copy how Undertale dialogue works.

To accomplish this, I need to find a way to get them to actually speak specific words in the first place. There are two ways I found. The first one is concatenative synthesis, in which you effectively stretch and stitch pre-recorded phonemes together to create speech.

A lot of complications emerge when taking into account things like actually interpolating and resampling these audio clips, so for now, I'm looking into the second method I know of.

The Source-filter model of speech synthesis was created by a few people much smarter than I many years ago. I won't try to give an entire thesis on its more fine details, for reasons including the possibility that I might be dead wrong on a few points. Instead, I'll give some generalizations to minimize any wrong explanation.

The basis of this model involves creating objective measures of how someone's voice should sound based on the current configuration of their vocal tract and mouth, which can be simplified to a source waveform, modified by a filter.

This filter mimics the resonant frequencies of someone's vocal tract depending on the vowel or consonant being pronounced. These frequencies are called the formants, or frequencies that form a phoneme.

For example, here's a spectrogram of the main frequencies (formants) that make up the word "hide", and how they change over time:

Not counting the pause in between, you can see how short the "de" in hide is actually stressed. Most of the other things that I learned I'm a lot less certain of, like onset time, pauses involving consonants, and the periodicity and fine structure of the source waveform.

So that's what I did this week, just researching way to synthesize speech in a somewhat customizable way.