Building 6-Minute Podcast Documentaries with Sponsor Filtering and Loudness Normalization
The PastNoise team describes a deterministic pipeline that assembles quotes from many podcasts into roughly six-minute, broadcast-ready documentary reels in the browser. Rather than using costly language-model classifiers, the system relies on structural heuristics such as hashing recurring sponsor copy across episodes, which the authors say catches over 80% of host-read ads. Utterances near detected ad transitions are buffered, and adjacent speaker turns are kept together up to natural pauses. To address widely varying recording levels, the pipeline measures integrated loudness, true peak, and loudness range per segment and applies two-pass normalization to -16 LUFS with a -1.5 dBTP ceiling and crossfades between channels. The write-up also points readers to open bilingual topic datasets and a related quote-finder tool for language learners.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in