Front-load the free work — turning a 90-minute recording into short clips
The obvious way to cut a 90-minute recording into short clips is to hand the whole thing to a model and ask it to find the good parts. I tried that. It's expensive and the output degrades as it goes. Why the obvious way fails 90 minutes of transcript is an enormous amount of context, and long context degrades judgement . The model starts confusing similar passages, loses track of what it already…
The most efficient way to transform a 90-minute recording into numerous short clips is to first process the material locally, before feeding it to a model. This approach proved to be much more cost-effective than sending the entire recording to a model, as the model's output diverges quickly due to complex context. After transcribing the recording and converting the word-level transcript to phrase-level markdown, the model can analyze the condensed content in a single pass, significantly reducing costs.
The pipeline consists of five steps: transcription, phrase-level packing, candidate segment selection, scene detection, and silencing. The most expensive step, where human judgment is required, is selecting relevant segments based on topic. Before rendering the final clips, it is essential to save the selection details, including timestamps and reasons, in a file for future reference.
This checkpointing technique ensures that rendering can be repeated cheaply, allowing for adjustments such as aspect ratios without having to restart the process. The pipeline is designed for iterative use, making it more efficient and adaptable to changes.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
