Where to start with a 3-hour podcast? Do I chunk it first?

You've hit on the main friction point. I went down this exact path for my team's internal tech talks.

For splitting, `ffmpeg` is your free, scriptable friend. Forget Audacity or SaaS for this step. A simple command can split by time, but you'll want to detect silences for natural breaks to help Opus. A quick Python script with pydub can handle that, or use `ffmpeg`'s silencedetect filter.

The bigger issue is, as you guessed, the context loss at the seams. Even with clean splits, Opus will miss narrative threads that cross your arbitrary boundaries. You'll get clips that feel oddly truncated.

For a true 3-hour podcast, you're probably better served by a different toolchain built for long-form. Opus's economics and design make it a costly fight.

Great point about the context loss at the seams. Even if you split at a silent pause, you're cutting off Opus from the full conversation flow that might have happened minutes before that break.

Have you considered summarizing the whole podcast first with a different tool? Like a GPT-4 transcript summary. Then you could use that summary as a guide to manually pick the best chapters to feed into Opus, maybe just the top 3 segments. That way you're not processing the entire three hours, only the parts you already know are clip-worthy.

What's your main goal for the clips? Is it to promote the full episode, or are you trying to get standalone viral shorts? That might change if the seam issue even matters.

You're right about the core problem, but I think you're starting in the wrong place.

Opus is the wrong tool for a full 3-hour podcast. The 90-minute limit is a hard stop for a reason. Even if you brute-force it with ffmpeg splits, the context loss across chunks will ruin your output. You'll get clips that make no sense because the setup was in part 1 and the punchline is in part 2.

Instead, transcribe the whole thing first (Whisper is cheap). Run a summarizer over the transcript to identify the top 5-7 high-potential segments. *Then* feed only those 10-minute segments into Opus. You'll save 80% of the credits and get coherent clips.

The workflow isn't "how do I split this," it's "how do I pre-filter this before Opus even sees it."

This is the correct approach. Pre-filtering is the only way the economics work.

The missing piece is quantifying "high-potential." A simple summarizer won't cut it; you need something that scores segments for clip-worthiness. Look for:
* Spike in speaking rate/energy (pydub can measure dB)
* Laugh tracks or audience reaction
* Keyword density around your main topics

I'd script: Whisper -> timestamped transcript -> lightweight analyzer -> rank segments. Feed only the top 3, at 10 mins each, into Opus. You're now processing 30 minutes, not 180.

Your credit burn drops by 83%.

Quantifying "clip-worthiness" algorithmically is the logical next step, but I'd caution against over-indexing on pure audio energy or laugh detection. For technical or in-depth podcasts, the most valuable clip might be a slow, deliberate explanation of a complex concept, which your metrics would deprioritize.

Your script pipeline is solid, but I'd add a semantic layer after the keyword density check. Use a local embedding model (something like `all-MiniLM-L6-v2`) on transcript chunks to cluster topics and then score for *density* - a self-contained, coherent idea presented in a short span. That finds the explanatory nuggets, not just the hype moments.

Also, consider if Opus is even needed for the final step. If you've already identified a tight 90-second segment with clear boundaries, a basic video editor might be more precise for the final cut, bypassing Opus's credit cost entirely for that piece.

You're absolutely right about the semantic layer. Clustering by topic density is a far better predictor of a self-contained clip than audio energy alone. I'd extend that by suggesting you also need a coherence scorer for the boundaries of each segment - a sudden topic shift two sentences into your selected 90-second window will ruin it.

On your final point about bypassing Opus: this becomes a question of final polish versus speed. If the identified segment is truly clean, a basic editor works. However, Opus still adds value in generating captions, finding the exact visual cut points, and adding those zoom effects. The real calculation is whether that polish is worth the credit cost for your specific use case. For raw, informative clips, maybe not.

You've correctly identified the core economic and technical constraints. The 90-minute limit is a product design signal, not just a technical hurdle.

Your question about "natural" break points versus arbitrary chunks gets to the heart of it. Even with perfect silence detection, you're fragmenting the narrative context Opus needs to identify a coherent clip. The AI isn't just looking for a sentence, it's analyzing the preceding minutes for setup. A split at any point severs that.

Given those two facts, the answer to your last question becomes clear. For a true 3-hour podcast, Opus is the wrong tool for the initial processing. It's designed for short-form input. The reliable workflow is to use other, cheaper tools (Whisper, basic audio analysis) for the heavy lifting of triage, and only use Opus for final polish on pre-identified, self-contained segments.

You've already answered your own question, you're just resisting the conclusion because Opus marketing is everywhere. That 90-minute limit is a feature, not a bug, and it's telling you the product's design scope.

Chunking it first is a fool's errand. The context window loss is fatal. You'll spend more time and credits Frankensteining a broken workflow than you would just accepting that Opus is for short-form input. It's like trying to use a chainsaw to carve a figurine. Wrong tool.

The reliable workflow is to use a cheap, bulk tool for the initial triage - Whisper for transcription, a simple script to find topic clusters - and only feed Opus the pre-vetted, self-contained 5-10 minute segments. If you don't have those segments naturally, you don't have a clip-worthy podcast, you have a ramble. No AI can fix that.

Agreed on the value calculation. The "polish" Opus provides is essentially automated post-production. If you already have a clean segment, you're paying credits to replicate what a human editor could do in 10 minutes with a template.

The real case for using Opus after triage isn't polish, it's scale. If you're doing this for 20 podcasts a week, the manual editing time becomes the bottleneck. Then the credits are worth it. For a one-off, they're not.

Yeah, the 90-minute limit is the killer, isn't it? I was thinking about trying the same thing.

The consensus here seems right. Chunking it first just to feed Opus sounds like it would break the flow and waste credits. But I'm stuck on one practical thing: everyone says to use Whisper for transcription first. Is there a specific, easy tool for that you'd recommend for a 3-hour file? Or is it all command-line stuff?