Audio Speaker Merger (AIIA)

The Speaker Isolator pulls voices apart; this node puts them back together. It mixes two audio streams into one, and its reason to exist is the multi-speaker talking-head workflow: you isolate speaker A and speaker B, drive a video node with each voice separately, and then you need to recombine their audio tracks so the final video has the full conversation. That "separate, process, recombine" cycle is the backbone of the AIIA podcast/dialogue pipeline, and the merger is the last step.

How it works

It takes two AUDIO tensors and mixdowns them (waveforms summed). The fiddly part - and the reason there are five modes - is that the two tracks won't be the same length, and the merger needs to decide what the output length is. Everything else is about not blowing up when you sum two loud signals.

The inputs

• audio_1, audio_2 - the two streams. Order matters mostly for the length modes below.
• duration_mode - how the output length is decided:
  • Longest (default) - output runs as long as the longer input. The usual choice for recombining two speakers' takes, because nobody wants the tail of the conversation cut.
  • Shortest - run to the shorter input; useful when you're layering and want a guaranteed trim.
  • Audio 1 / Audio 2 - pin the output to one input's exact length.
  • Specified - you name a length in seconds via specified_duration.
• specified_duration - active only in Specified mode, default 10s. Ignore it otherwise.
• normalize - on by default. The tooltip is the whole story: it automatically normalizes levels so the summed audio doesn't clip. When two loud tracks add up, 1+1 can exceed 1 - this is the safety.

Output

One output: merged_audio (AUDIO). Straight into a save node, a video combine, or another merge if you're stacking more than two tracks (it's a binary mix, so N speakers means N−1 of these).

Where it fits

The canonical round-trip is Speaker Isolator (Maintain Duration) → [per-speaker processing or video driving] → Speaker Merger, then into Audio Post-Process for the final normalize/mastering pass. If you kept Maintain Duration on the isolators, both tracks are the same length as the original and Longest just works.

Gotchas

• Summing clips when normalize is off. If you toggle it off, watch the level meter - two normalized speakers stacked will clip hard on peaks.
• Length mode confusion. If your merged audio is mysteriously short, check whether Shortest got selected, because Specified's 10-second default also hides there waiting to bite someone who sets the mode but not the value.
• It's a dumb mix, not a smart remix - no panning, no ducking. For a real "one speaker talking over another" effect you'll do better driving one speaker at a time and using silences, which is what the rest of the pack assumes anyway.