ChatTTS Text Splitter

Feed ChatTTS a whole article in one go and quality drops, VRAM spikes, and long sentences start sounding strained. The Splitter is the fix: it cuts your text into chunks the sampler can chew on, sized to fit your GPU. It's a preprocessing node with a single job - and a slightly surprising way of doing it.

How it works

You give it text and a split_method, it returns a STRING - one string, not a list. That's the part that catches people off guard. The splitter doesn't hand you separate chunks to process individually; it splits your text and then joins it back together with newlines. The reason: the Sampler's split_batch setting (anything above 0) is what actually tells ChatTTS to treat newline-separated lines as a batch. So this node and split_batch are a pair - the splitter prepares the text, the sampler executes the batching.

The settings

• text - required, multiline. Usually the output of a Text Normalizer if you have one upstream, or raw script text.
• split_method - pick the granularity:
  • newline (default) - splits on the line breaks you already wrote. Predictable and honest: you control the chunk boundaries.
  • sentences - splits after ., !, and ?. Good for prose where paragraphs are too long.
  • paragraphs - splits on blank lines. Coarsest option; fine if your paragraphs are short.
  • none - leaves the text untouched. There to make the node pass-through when you want the same graph shape everywhere.
• max_batch_size (1–32, default 4) - the cap on how many chunks survive. If your text splits into more than this, the extras get dropped. So it's a limit on both batch size and how much of your text actually gets spoken. Keep it in line with what your GPU handled before.

Where it fits

Text (Normalizer →) TextSplitter → ChatTTS_Sampler (split_batch: 4)

Set the splitter's max_batch_size and the sampler's split_batch to the same number and you have a long-form pipeline that's reproducible. For short one-liners, skip it entirely - it's overhead. For anything over a couple of paragraphs, it's the difference between a smooth read and a garbled one.

The one trap to remember: because oversized chunks are silently truncated, a missing sentence at the end of a long script usually means your max_batch_size was too small, not that the model "ate" your text. Nudge the number up and rerun.