💬 AIIA Text Splitter

The AIIA podcast pipeline was built around scripts: Speaker: line pairs that a parser turns into structured data, one chunk per line. That's great for two-voice dialogues. It's useless for a single-narrator voiceover, where your "script" is just a wall of prose. This node is the adapter: it takes plain long-form text and splits it into the same dialogue_json structure the script parser produces, so a narrator can flow through the pack's Emotion Annotator → TTS chain exactly like a scripted character.

The use case that makes it shine: audiobooks, YouTube voiceovers, long single-voice narration where you want per-sentence emotion tagging later. You don't hand-write Narrator: ... for every line - you paste the whole chapter and let this node do the bookkeeping.

How it works

It's a sentence splitter with judgment. In the default auto mode it does four things in order: keeps paragraph breaks as structure; splits within each paragraph on sentence-ending punctuation (。！？!?…, including ellipsis); merges any sentence shorter than min_chars into the previous one (so fragments don't become two-second audio blips); and if a sentence runs past max_chars, it re-cuts it at a comma or semicolon. The result is a list of chunks that a TTS can actually speak naturally, each stamped with your speaker_name.

The other two modes are simpler: by_sentence only splits on terminal punctuation (no merging, no re-splitting), and by_line splits on newlines only - useful when your text is already one-sentence-per-line.

The inputs that matter

• text - the long-form text. Multiline; paste it all.
• speaker_name - default Narrator. This is what gets written into every segment's speaker field, so pick something your downstream nodes expect.
• split_mode - auto is the right default for most prose. Reach for by_line when your input is already chunked.
• min_chars (default 4) / max_chars (default 100) - the two tuning knobs. If a short line keeps vanishing into the previous one, raise min_chars... no, lower it. If long sentences keep getting chopped mid-thought, raise max_chars.

Outputs: dialogue_json (the STRING you wire into Emotion Annotator or a TTS node) and sentence_count (an INT, handy for sanity-checking the split or driving a loop).

Install

Standard pack install - this one has no extra dependencies at all:

cd ComfyUI/custom_nodes
git clone https://github.com/havvk/ComfyUI_AIIA.git

or ComfyUI Manager → search "ComfyUI_AIIA", then restart.

Common issues

• A sentence got cut in the wrong place. That's max_chars doing its job too eagerly - nudge it up, or switch to by_sentence and split the long line yourself.
• Fragments all over the dialogue JSON. The min_chars merge threshold is too low; the node is faithfully keeping every two-word fragment as its own line. Raise it.
• The Chinese/English mix. auto handles both, but the character-count thresholds behave differently across scripts - a 100-char Chinese sentence is a lot more speech than 100 English characters. Expect to tune.