Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

My TTS read a whole novel. The model was the most reliable part of the pipeline

Description and use

In a Crimean Tatar TTS project, preprocessing bugs caused high error rates. Removing sentence gluing, refining breath trimming, and adjusting pauses reduced errors to 1.84 %.

Problem

The TTS pipeline produced errors due to text and audio preprocessing steps rather than the model itself.

Outcome

After correcting preprocessing bugs, the TTS pipeline produced a 1 hour 58 minute audio of a 16‑chapter Crimean Tatar novel with a character error rate of 1.84%, demonstrating that model quality was high and issues lay in text handling.

Workflow

  • 1. Verify text is not in training data (zero 6‑word matches).
  • 2. Render full chapter without gluing short sentences.
  • 3. Benchmark post‑processing steps (gluing, breath trimming, pause insertion) against doing nothing.
  • 4. Replace breath trimmer with loudness‑based trimming.
  • 5. Adjust pause durations based on human narrator statistics.
  • 6. Re‑render chapters and compare character error rates.
  • 7. Run extraction and guard checks on full book to catch formatting bugs.

Evidence

The model held up. The problems weren't in the model but in the text and audio preparation around it

Source

On glued chunks the model started throwing out entire sentences

Source

The chapter scored better with no trimming at all than with my "fix"

Source

  • Record date: 2026-10-09T18:55:51.397Z