AIIA Dialogue TTS (Multi-Role)

This is the multi-role endgame node of the AIIA pack: it takes a parsed dialogue script, assigns each character a voice, runs the whole thing through a TTS engine, and stitches it into a single conversation track - with pauses, turn-taking, and per-speaker emotion preserved. If you've ever manually wired three TTS nodes and mixed the results in an audio editor, you know why this exists. It's the "podcast in a node" step, and it's the reason the pack ships a script parser and a stitcher around it.

The pipeline it belongs to

Script Parser → [Emotion Annotator] → Dialogue TTS → Podcast Stitcher. This node expects a dialogue_json - the structured script from the pack's AIIA Podcast Script Parser - in the 角色名: 台词 (Character: line) format. Feed it raw text and it won't know what to do; the parser is the required upstream. The optional Emotion Annotator (an LLM that tags lines with [Happy]/[Sad]-style emotions) slots in between, and this node reads those tags if they're present.

The inputs

• dialogue_json - the parsed script. This is the load-bearing input.
• tts_engine - CosyVoice (default), VibeVoice, or Qwen3-TTS. Pick the engine whose model you connected; CosyVoice is the default because it handles emotion tags and dialects natively.
• pause_duration - seconds of silence between lines, 0.5 default. The (Pause N) tags in the script override this per-line.
• speed_global - a master tempo, 1.0.
• batch_mode - Natural (Hybrid) (default: groups lines sensibly for the engine), Strict (Per-Speaker) (each speaker's lines batched together - best when a voice model keeps consistency per speaker), or Whole (Single Batch).
• speaker_A/B/C ref/id/emotion/dialect - up to three speakers, each with an optional reference audio (_ref), an optional built-in id (_id), a per-speaker emotion default, and a dialect. The _ref is your zero-shot clone source; _id is for engines with fixed voices.
• Model slots - cosyvoice_model / vibevoice_model / qwen_model; connect the one matching your tts_engine.
• Sampler panel - max_batch_char (1000), cfg_scale (1.5), temperature, top_k, top_p; passed through to the underlying engine when relevant.

Outputs

• full_audio (AUDIO) - the whole conversation, mixed, with pauses.
• segments_info (STRING) - per-line timing/segment metadata, which the downstream stitcher and subtitle nodes consume.

Where it fits and honest caveats

This node is the middle of a bigger machine, not a standalone. The full flow - parse → annotate → dialogue TTS → stitcher → subtitle/web-export - is the pack's flagship "generate a podcast from a script" story, and it's genuinely impressive when it works. The honest caveats:

• It's engine-dependent by design. CosyVoice gives you emotion tags and dialects; VibeVoice gives you long-form naturalness; Qwen3 gives you instruction control. Your results change with the engine, so test the same script across them.
• Speaker consistency is the real job. If Speaker A keeps drifting into Speaker B's voice across a long script, switch to Strict (Per-Speaker) batching - that's exactly what the mode is for.
• Reference audio quality sets the ceiling. A noisy 10-second clip of a voice gives you a noisy cloned voice for the whole conversation.

Troubleshooting

• No output / wrong engine error - you picked CosyVoice in the dropdown but connected a qwen_model, or vice versa. Match them.
• Lines run together - raise pause_duration or add (Pause 1.0) tags in the script.
• Emotions ignored - the engine must support them (CosyVoice does; VibeVoice explicitly doesn't, per the source) and the tags must come from the annotator or your script.