Conversation

Notices

1. Embed this notice
   alcinnz (alcinnz@floss.social)'s status on Wednesday, 29-Jan-2025 04:16:31 JST alcinnz

   Once the computer figures out how to pronounce some text, how do we convert those "phonemes" into audio our speakers (& in turn our ears) understand? How'd we do so on our Lua machine?

   First we need to enforce some emphasis rules, & interpret those phonemes to determine how to shape our computer's simulated vocal tract to produce those noises.

   The voice you've selected (they're small, we can fit multiple) can tweak this vocal tract & pronunciation rules.

   1/3?

   • Embed this notice
     alcinnz (alcinnz@floss.social)'s status on Wednesday, 29-Jan-2025 04:21:15 JST alcinnz

     in reply to

     I'm sure we can do better today even without ML, I'm sure our hypothetical Lua-Machine can do better... Which I understand @patchlore is working on...

     But I'll describe the iconic model of a vocal tract a rushed Dennis Klatt implemented in the 1980s for a then-recently paralysed Prof. Stephen Hawking. Since that's what I'm familiar with!

     The primary component of which is a "resonant" which is a multiply-sum over its 3 most-recent inputs. We stack a whole bunch of these!

     2/3!

   • Embed this notice
     alcinnz (alcinnz@floss.social)'s status on Wednesday, 29-Jan-2025 04:38:57 JST alcinnz

     in reply to

     At the base of this stack of resonators is some input audio, which may be:
     * raw samples of a vowel sound,
     * a special waveform (vWave, a.k.a. glotal flow wavefrom),
     * or some hardcoded inputs to an initial resonator.
     Into which we multiply-in breathing, represented by random noise shaped by a sine-wave.

     We'll probably want to add additional sound-effects (e.g. subtle echoes), but that's the core logic! For making this Lua machine useful to the blind!

     3/3 Fin for today!