Behind the Scenes

Inside PodMagi: How the Editing Engine Works

A behind-the-scenes look at the multimodal AI that analyzes audio cadence, cuts, and visuals to extract a video's editing signature.

7 min readBy the PodMagi team

A look at what happens between upload and export, written for people who want to know whether the claims hold up rather than for people who want a marketing summary.

Three passes over the audio

The audio track is analyzed before anything touches the video, because audio carries most of the structural information. The first pass is transcription with word-level timestamps. The second measures prosody — pitch, amplitude and pacing — which is what identifies emphasis. The third detects speaker changes.

Together these produce a timeline annotated with what was said, when, by whom, and with how much stress. Nearly every subsequent decision reads from this timeline.

Visual analysis of the reference

In parallel, the reference video is measured for the parameters described in our style-cloning piece: shot-length distribution, scale-change behaviour, caption geometry, colour treatment, and the ratio of speaker-time to cutaway-time.

The output of this stage is a compact profile — a few dozen numbers. It contains no imagery from the reference video and cannot reconstruct it. This is a deliberate design constraint, not an incidental property.

Composition and render

Segment selection runs against the annotated transcript. Each selected segment is then cut according to the profile, reframed using face and motion tracking to keep the subject in the vertical crop, captioned at word level, and given B-roll at transcript-anchored positions.

Rendering is compute-heavy and runs in parallel across segments, which is why a sixty-minute source producing twenty clips does not take twenty times as long as one clip.

Where it fails

Overlapping speech degrades speaker separation and therefore reframing. Heavy background music mixed into the source competes with prosody analysis. Strong accents can reduce transcription accuracy, which propagates into caption and B-roll timing. Source footage under roughly 720p limits useful vertical cropping, since the crop discards pixels.

We list these because a tool that claims no failure modes is describing a demo rather than a product.

Try it on your own footage

Thirty processing minutes a month on the free plan. No credit card required.

Get early access

Keep reading

← All articles