Skip to content

Lessons & what’s next

Read the model’s source

The single biggest fix (the 6 s / 10 s conditioning window) was a constant in the model code, not in any documentation.

The pipeline matters as much as the model

Silence, pauses, loudness and chunk boundaries decided whether the voice sounded human.

Measure, then decide

Mixed precision “should” have been faster; it wasn’t. The best-scoring setting “should” have won; it babbled.

Small, reversible changes

Every phase was tested and A/B’d before the next one, with the test suite green throughout.

  • Mixed precision on a 6 GB laptop GPU: slower and hungrier, so removed.
  • Higher expressiveness (≥ 0.65): higher similarity on paper, unusable babbling in practice.
  • Free hosting: no free tier can run GPU inference; see why it isn’t online 24/7.
  • A cleaner, longer reference recording: the largest remaining quality lever.
  • Reading the real client IP from the tunnel header so per-visitor limits work when shared.
  • Exploring a 2.5D cartoon video pipeline that reuses these narrations for short-form video.