跳到正文
DEV Community· nonacosa·· 1 小時前AI 評分37

PastNoise 如何將跨 Podcast 引文編成經 -16 LUFS 音量標準化的 6 分鐘紀錄短片

Compiling 6-Minute Cross-Channel Podcast Documentaries with -16 LUFS Normalization

AI 導讀

PastNoise 建立確定性流程,將不同 Podcast 的主要發言引文剪接成 6 分鐘紀錄短片,並在瀏覽器中製作。流程以結構啟發式過濾廣告,再將音訊標準化至 -16 LUFS;其資料涵蓋 207 million spoken lines 和 2,494+ channels。作品亦可在 PastNoise Discover 瀏覽,並提供開放的雙語主題資料集。

正文

nonacosa

When you search across thousands of podcast episodes for a single thesis, finding the quote is only half the work. Listening to ten disconnected 20-second clips from ten different shows with wildly different microphone gain quickly causes listener fatigue.

While building PastNoise across 207 million spoken lines and 2,494+ channels, we needed a deterministic pipeline that stitches primary-voice quotes from different podcasts into cohesive, broadcast-ready 6-minute documentary reels directly in the browser.

1. Filtering Sponsor Reads Before Quote Assembly

Audio compilation breaks down immediately if ad reads sneak into thematic reels. Instead of running expensive LLM classifiers over every segment, we use structural heuristics before indexing:

  1. Cross-episode boilerplate hashing: Sponsor copy often repeats verbatim across dozens of episodes from the same RSS feed. Detecting recurring n-gram windows catches over 80% of host-read ads deterministically.
  2. Timestamp boundary clamping: Utterances within 4 seconds of a detected ad transition are buffered so partial sponsor intros never leak into adjacent documentary segments.
  3. Speaker-turn continuity: When assembling multi-minute answers, we preserve adjacent speaker turns up to natural pause boundaries instead of slicing mid-thought.

You can browse these sponsor-clean dossiers across 8 knowledge pillars on PastNoise Discover or explore all 146+ spoken knowledge topics.

2. Normalizing Heterogeneous RSS Audio to -16 LUFS

Podcast feeds vary dramatically: studio interviews often master around -16 LUFS, while Zoom calls or archival lectures sit anywhere from -28 LUFS to -11 LUFS.

Our preparation pipeline measures integrated loudness (I), true peak (TP), and loudness range (LRA) per segment, applying two-pass loudnorm targeting -16 LUFS with -1.5 dBTP ceiling and a 350ms crossfade envelope between channel switches.

Details on our ingestion and normalization architecture are documented on the PastNoise Methodology page and in the PastNoise Chinese portal.

3. Open Data & Sentence-Level Sister Tools

來源:DEV Community · dev.to