Kokoro invents a mid-sentence pause ("gonna ..... one") in a comma-free line

[genwave-radio]

🎧 What aired

Demo, 2026-09-26 22:40:23Z, Maxxie Volt lead-in for "Sunshine" (af_heart, pace 1.15). The script line:

I'm already feeling like it's gonna be one of those days where anything goes and you can't help but smile

aired as "it's gonna ..... one of those days": a clear pause after "gonna", and "be" nearly swallowed. There's no comma, and GenWave adds no markup inside the sentence.

🔍 Root cause: a Kokoro model artifact, not our markup

• Neither the persona corrections (none) nor the demo Tts__Corrections match any word in the line.
• KokoroSpeechMarkup adds only the 0.6 s sentence pause after "right up!".
• Re-rendered on local kokoro-fastapi-cpu:v0.6.0 (the same pin): it's deterministic. The render is 17.54 s, the same length as the one that aired, and three runs came out identical.
• /dev/captioned_speech word timings for the sentence on its own:
  • gonna@1.15-1.64 (0.49 s), with a 0.25 s silence at 1.49–1.74.
  • be@1.64-1.80 (0.16 s).
• Controls:
  • "We're gonna be here all night" → gonna 0.27 s, no gap. The word isn't the trigger; the problem depends on the sentence.
  • Pace 1.0 → still 0.58 s. Pace isn't the trigger.
  • "going to" → no gap.
  • The phonemes are normal: ɪts ɡˈʌnə bi wˈʌn.

🛠️ Options

1. Check a newer kokoro-fastapi or model when the pin next moves; the artifact may be gone.
2. Detect it after render: an internal silence of 0.2 s or more that isn't at a pause tag. Then re-render, e.g. with a small speed nudge, or accept.
3. Station-level word fix gonna → going to. It's a blunt fix and costs casual tone.

No action is needed for launch; it's rare and depends on the exact wording.