@tspivey @jackf723 I’ll need to tune it more after work, but I did track down and fix a real bug in this hotfix. At speech rates above 70, diphthong glides were losing their sweep entirely.
Root cause: diphthongs are generated as a sequence of micro-frames that interpolate formant frequencies from onset to offset. Each micro-frame has a minimum duration (two full pitch periods) so the cascade resonators can properly settle. When speech rate increases, the overall diphthong duration compresses. That caused the micro-frame count to drop from 4–5 down to just 3. With the onset hold curve packing those three points toward the start, the result was a single large formant jump instead of a smooth transition. The glide effectively collapsed into the onset vowel and sounded flat.
The fix was to raise the minimum micro-frame count from 3 to 5. That guarantees enough interior waypoints to preserve a smooth sweep regardless of speech rate.
I also retuned the MOUTH onset vowel. It was sitting too close to LOT in formant space, so “found” was drifting toward “fond.” I increased the separation, so “sound” and “found” should now have more perceptible rounding in the glide compared to beta 3.
One remaining issue is amplitude. It can still pop out of the word center more than other segments — subtle in “thousand,” but obvious mid-utterance in “sound” or “found.” There’s likely still a gain transition to smooth out.
Other platform releases will include this fix as well soon.