MOSSTTSKit normalizes user text before tokenization and chunking. The goal is to turn app-facing text into a safer TTS prompt while preserving the intended reading content.
MOSS-TTS-Nano is sensitive to some punctuation patterns. Raw book text, copied web text, and speaker labels can contain punctuation that looks natural to a reader but behaves like an unfinished or unusual prompt to the model.
The normalization layer handles those cases in one place instead of scattering small fixes across synthesis, chunking, and app integrations.
- Leading and trailing whitespace is removed.
- Chinese quotation marks such as
“”,‘’,「」, and『』are removed. - Chinese and ASCII ellipses are converted to sentence boundaries.
- Repeated dash separators such as
---,--, and——are converted to sentence boundaries. Single hyphens in words such asMOSS-TTS-Nanoare kept. - Non-empty line breaks are treated as phrase or sentence boundaries.
- A dangling final clause punctuation mark, such as
：,:,，, or,, is converted into a sentence terminator. - Repeated spaces introduced by normalization are collapsed.
Examples:
利娜正睡在我身边……
我一点都不想再睡了。
becomes:
利娜正睡在我身边。 我一点都不想再睡了。
Taiguanglin：
becomes:
Taiguanglin.
“米歇，准确地讲，是在一百三十五万年以前。”
becomes:
米歇，准确地讲，是在一百三十五万年以前。
非常值得一读，---揭示地球史前文明。
becomes:
非常值得一读。 揭示地球史前文明。
New text preprocessing behavior should be added through TextNormalizer, with a
focused regression test in TextNormalizerTests. If the behavior affects real
model output, also add or update an integration test or listening sample that
uses the original problematic text.