🧵 AIIA Podcast Stitcher

Generating two voices separately is easy. Fusing them back into one natural-sounding conversation is where multi-voice podcast workflows actually die - naive concatenation gives you clipped syllables, botched timing, and dialogues that sound like two robots passing the mic. 🧵 AIIA Podcast Stitcher is the reassembly node, and it's the most technically interesting part of the split-stitch pipeline: it uses speech-recognition timestamps to find where each sentence actually starts and ends in the audio, then interleaves the tracks in the original script order.

How it works

The inputs are split_map (from AIIA Podcast Splitter, telling it the original order), the two separately generated tracks audio_A and audio_B, and - here's the key - asr_A and asr_B, the ASR results for each track with word-level timestamps. By matching the script sentences against the ASR word timeline (via exact match, then edit-distance fuzzy matching to tolerate ASR errors), it locates every sentence's true boundaries and slices exactly there. Then it interleaves A/B per the script order with a cosine fade at each join so the handoff is inaudible.

The inputs that matter

• asr_A / asr_B - required, and easy to miss. These are the ASR_RESULT objects from the pack's AIIA ASR Node (FunASR-based, produces word-level timestamps). Without them there's no boundary detection. This is the #1 wiring mistake in this workflow - people hook up the audio and forget ASR, then wonder why the stitcher can't find anything.
• gap_duration (0.25s default) - silence inserted at speaker handoffs. The natural-conversation dial; raise it for a more deliberate back-and-forth.
• padding (0.1s default) - breathing room kept before/after each slice so you don't clip the start of syllables.
• fade_ms (30ms default) - cosine fade length at slice edges. Longer is smoother but can smear consonants.
• use_vad - optional Silero VAD (auto-downloads ~2MB on first use) for precise speech-boundary detection. Good when ASR timestamps are jittery.
• use_forced_align - the high-precision option: Facebook MMS forced alignment (~1.2GB model) does character-level alignment, with Chinese text auto-converted to pinyin. Slowest and most accurate; the README describes a hybrid strategy that uses FA for onsets and energy detection for natural endings.

Outputs are the stitched audio and a segments_info JSON (timeline per line), which feeds AIIA Subtitle Gen - so you get a subtitled, speaker-tagged result straight out of the stitcher.

When you'd use it

Only in the split-stitch flow - per-speaker TTS, then reassembly. If you're using the pack's single dialogue node (CosyVoice or Qwen3), you don't need it, since those already output a combined timeline. But for the "different engine per character" setup, this is the node that makes the result sound like a real conversation instead of a cut-and-paste job.

Install

Standard pack install: havvk/ComfyUI_AIIA via ComfyUI Manager or git clone https://github.com/havvk/ComfyUI_AIIA.git into custom_nodes/, restart. The Silero VAD and MMS FA models auto-download on first use; the forced-align one is a big download, so only flip that toggle when you need the precision.