Audio Sync

The Audio Sync node measures how far apart the clocks of 2 to 6 recordings of one conversation are, from their sound. A podcast is often recorded several times over: one camera per guest, a wide shot and a main microphone, each started at its own moment. Audio Sync listens to all of them and reports how far each recording is shifted against one reference recording. A multicam edit then lines up without anyone typing offsets.

The node runs locally, with no AI model and no provider key, so it works on every edition.

When to use it

• Line up a multicam podcast, with one camera per guest and a main microphone, before you plan the edit.
• Find the offset of a separate audio recorder, such as a recorder on the table, against the camera sound.
• Check whether two recorders drift apart over a long session.

Quick start

Input and output

Each recording is identified by the node it comes from. That node's id becomes the recording's sourceId in the result, and it is also the id an edit plan gives the same recording.

The output, Offsets, is data, not audio. Wire it into the Offsets (Audio Sync) input of Edit Plan, which lines up a multicam edit with it, or into another node that reads data, such as Extract Field.

Settings

How it measures

The node compares each recording with the reference in three passes:

1. Coarse. It compares the loudness of the two whole recordings. This finds the shift at any distance, even for a camera started minutes before the microphone.
2. Fine. It compares the waveforms around that first answer, on up to three 10-second windows of the stretch that both recordings share. Each frequency is weighed by its phase, so a camera across a room with echo is timed on the direct sound, not on the echo. Clean recordings land within about 1 millisecond, and recordings with echo usually within a few.
3. Drift. Separate recorders' clocks run at slightly different speeds, commonly 20 to 100 parts per million. The node fits the windows' offsets to a line, and the slope of that line is the drift. It measures drift up to 200 parts per million.

Times are read on each file's own clock, the clock that edits are cut on. A video whose sound starts after its picture is measured with that gap in place.

Read the result

The Offsets output looks like this:

{
  "version": 1,
  "reference": "mic",
  "offsets": [
    { "sourceId": "mic", "offsetMs": 0, "confidence": 1, "driftMsPerHour": 0 },
    { "sourceId": "camA", "offsetMs": 7500, "confidence": 0.94, "driftMsPerHour": 1.2 },
    { "sourceId": "camB", "offsetMs": -40000, "confidence": 0.91, "driftMsPerHour": null }
  ],
  "notes": []
}

The offset rule

Reference time = recording time + offset. This is the same sign rule an edit decision list uses. When the reference is your main recording, each offset is exactly the value that recording takes in the edit.

• A camera started 7.5 seconds after the microphone has an offset of +7500: its second 0 is the microphone's second 7.5.
• A camera started 40 seconds before the microphone has an offset of -40000.

Confidence

• With a shared stretch of 6 minutes or more, the node measures three windows. When they agree within a millisecond, the offset is trusted, with a confidence of at least 0.6, even for a camera across a room with echo. A sharp match raises it to 1. Windows that disagree by 5 ms or more score 0.
• A shorter shared stretch has a single window, so both that match and the overall alignment must be clear.
• Low confidence is a result, not a failure. A camera whose microphone was off has nothing to align, and the run still completes.

Notes

Credits

Audio Sync costs 11 credits for each recording aligned to the reference. The reference itself is not aligned against anything, so it costs nothing. The length of the recordings does not change the price.

On the canvas, the price follows the number of recordings wired into Sources. An estimate made without the workflow's connections quotes the 6-recording price, 55 credits, so that it never quotes too little. Self-hosted Community and Business editions use no credits.

Examples

A microphone and one camera. The camera started 7.5 seconds after the microphone. The result is mic at 0, the reference, and cam at +7500. The run costs 11 credits.

A microphone and three cameras. Camera A started 7.5 seconds after the microphone, camera B 40 seconds before it, and camera C, recording at 44.1 kHz, 3.25 seconds after it. The result is mic at 0, camA at +7500, camB at -40000 and camC at +3250. The run costs 33 credits.

Six recordings, with camera A as the reference. The same offsets are measured on camera A's clock instead. Every value shifts by camera A's own offset, and camA is at 0. The run costs 55 credits.

Drift. Two recorders whose clocks differ by 100 parts per million move 360 ms apart per hour. Over a 40-minute shared stretch that is 240 ms, well over the 33 ms threshold. driftMsPerHour is about 360 in size, with a sign that says which way the recording slides. A note warns that the ends are off by up to 120 ms. Nothing is corrected.

Limits

• 2 to 6 recordings per run. With fewer or more, the run is refused before it starts, and no credits are held.
• Shared sound. Recordings must share at least 2 seconds of sound. A recording with none gets a note and a confidence of 0.
• The reference needs sound. It is the clock that everything is measured on. A reference without a sound track fails the run with a message that names it. Any other recording without sound only gets its own row with a confidence of 0.
• Drift is reported, not corrected.
• Long files. On its first run, a recording is fetched in full. The file can be up to 64 GB, and the download can take up to one hour. A file whose size is known must arrive within its size at 2 MB/s (16 Mbit/s), at least two minutes and at most one hour. After the first two minutes, the download must keep arriving at about 2 Mbit/s or faster. If it is slower than that for a whole minute, the fetch is abandoned. Later runs reuse the prepared audio.

Tips

• Make the main microphone the reference. Wire it first or choose it under Reference. The numbers are then easy to read: its offsets are exactly the values that the edit takes.
• Check low confidence by ear. It usually means the recording barely heard the conversation, such as a muted camera or a distant wide shot in a loud room.
• Read the drift notes on long episodes. More than a frame of drift means the lip sync slides over the episode, even when the start lines up.
• Line up the edit. Wire the Offsets output into the Offsets (Audio Sync) input of Edit Plan, with the same recordings wired into its Sources. Edit Plan moves each offset onto its own master's clock, so any recording can be the reference here. It stops before charging when a recording it uses was not measured or matched weakly. See Multicam recordings.
• Start from the Multicam Cut template. The Multicam Cut template in the template marketplace comes wired this way, with the master audio as the reference. It continues through Camera Switch to Apply EDL, which renders the cut.

From the API

POST /v1/audio-sync takes sources, a list of 2 to 6 objects with a unique id and a url, and an optional reference that must be one of the ids. A repeated id, or a reference that is not one of the ids, is refused with a 400 validation_error that names it. The finished run's output_data.json is the result above. The SDK method is client.edit.audioSync, the CLI command is nodaro edit audio-sync, and the MCP tool is audio_sync. See Voice and media.