Two inputs

A phone records the picture and a scratch audio track. A separate microphone often records a cleaner voice. --audio takes that microphone recording. talkinghead measures where it sits on the video timeline, remuxes the phone’s picture with the microphone audio into <run>/master.mov, and cuts that master. The phone’s own audio track is not used.

The measurement is written to align.json in the run directory:

• offset_sec: the video time at which the recording starts. Positive means the recording started late, so its head is padded with silence. Negative means it started early, so its head is trimmed.
• drift_ppm: the clock difference across the whole file. --fix-drift time-stretches the audio to cancel it, which matters for recordings longer than a few minutes.
• confidence: how clearly the winning match stood out. Below 3x the alignment is weak, and a run refuses to cut on it unless you pass --allow-bad-align.