The Podcast Production Stack That Runs Entirely on One MacBook

A friend who runs a two-person tech podcast told me his production ratio: one hour of recorded conversation costs him six hours of post-production. Editing, cleanup, transcription, show notes, titles, social posts. Six hours. That ratio is why most podcasts die before episode ten — not lack of ideas, lack of Saturday afternoons.

We rebuilt his pipeline on his MacBook Pro M3 (18 GB), and a 60-minute episode now ships in right around two hours of his attention. Nothing leaves the laptop except the final upload. Here’s the whole stack, tool by tool, with what each costs and the free alternative at every step — because the free path is genuinely viable, just slower.

Recording: capture clean, capture separate

The tool: Audio Hijack ($79, Rogue Amoeba). It records each source to its own file — your mic on one track, your remote guest’s call audio on another, with a backup recording running in parallel. Separate tracks are the single biggest gift you can give your future editing self: you can mute your keyboard clatter without touching the guest’s sentence.

The free alternative: QuickTime Player. File → New Audio Recording, pick your mic, record. It works, it’s already on your Mac, and for a solo show it’s honestly fine. The cost is one mixed track for calls and no backup recording. Middle option: Zoom records separate audio tracks for free (Settings → Recording → “Record a separate audio file for each participant”), which is how most remote shows on a budget should start.

One non-negotiable setting regardless of tool: record WAV or lossless, 48 kHz. Disk is cheap; a 60-minute WAV is about 600 MB. Every AI step downstream performs better on clean source.

Cleanup: where AI helps — and where it loses to a $99 microphone

Two AI cleanup moves earn their place:

Voice isolation. macOS has it built into FaceTime and the Voice Memos app, but for recorded files the practical options are the Voice Isolation effect in Final Cut Pro / Logic Pro, or Goodhertz style plugins. The free path: Audacity (free, open source) with its noise reduction — sample 2 seconds of room tone, apply, done. It removes the fridge hum and the air conditioning convincingly. My friend’s M3 processes a 60-minute file with Logic’s voice isolation in about 4 minutes.

Level normalization. Podcasts target -16 LUFS (stereo) / -19 LUFS (mono). The one-command free solution is ffmpeg:

brew install ffmpeg
ffmpeg -i episode.wav -af loudnorm=I=-16:TP=-1.5:LRA=11 episode_normalized.wav

That single filter does more for perceived quality than an hour of manual gain riding.

Now the honesty the title of this section promised: AI cleanup loses to a good microphone, every time. Voice isolation salvaging a laptop mic recording sounds like a salvaged laptop mic recording — slightly underwater, consonants softened. A Samson Q2U ($99) or Shure MV7 ($249) into a quiet-ish room needs almost no cleanup at all. Buy the mic before you buy any software on this list. AI is for rescuing the guest who joined from a hotel lobby, not for replacing the basics on your end.

Transcription: Whisper with timestamps is the keystone

Everything downstream — show notes, chapters, clips, editing — hangs off one artifact: a timestamped transcript. On Apple Silicon the tool is whisper.cpp (free) or its friendlier wrapper MacWhisper (free tier; Pro is a ~€59 one-time purchase for batch processing and speaker recognition).

The command-line path:

brew install whisper-cpp
whisper-cli -m ggml-large-v3-turbo.bin -f episode.wav -osrt -of episode

On the M3, large-v3-turbo transcribes a 60-minute episode in roughly 5–6 minutes; on my Mac Studio M2 Ultra it’s under 3. The -osrt flag gives you an SRT file with timestamps every few seconds. Accuracy on clear English speech is effectively human-level; my Czech-language test episodes come out around 95% usable with large-v3 — Czech podcasters, this works for you too, and Whisper handles the diacritics properly.

MacWhisper Pro adds speaker diarization (“Speaker 1 / Speaker 2”), which turns the transcript into something you can quote from directly in show notes.

The local LLM does the writing you hate

Feed the transcript to a local model via Ollama (ollama pull qwen2.5:14b — runs fine in 18 GB) and batch out every piece of derivative text in one sitting. The prompts that work, condensed:

Below is a timestamped podcast transcript.
1. Write a 150-word episode summary in the host's voice.
2. Propose 5 episode titles: 2 descriptive, 2 curiosity-driven,
   1 contrarian.
3. Generate chapter markers: timestamp + 3-6 word label for each
   topic shift. Use the SRT timestamps; do not invent times.
4. Extract 3 quotable moments under 200 characters each, with
   timestamps, suitable for social posts.

Set num_ctx to at least 16384 — a 60-minute transcript is ~10,000 tokens and the Ollama default will silently truncate it. The chapter markers paste straight into your podcast host as chapters; the quotes become the week’s social content. Time cost: about 10 minutes including your review pass. Before AI, this exact bundle — summary, title brainstorm, chapters, social copy — was a 90-minute writing session nobody enjoyed.

Editing: let the transcript drive the scissors

Two accelerators changed the edit itself.

Transcript-based editing. The concept: instead of scrubbing waveforms, you edit text and the audio follows. Descript ($24/month) is the famous one but it’s cloud-based; the local answer is Final Cut Pro’s transcript pane or the free-tier-friendly approach of using your SRT file as a map — read the transcript, mark cut points by timestamp, execute the cuts in GarageBand (free) or Logic Pro ($199 one-time). Reading is 5–10x faster than listening; you find the dead tangent in minute 23 by skimming, not by re-listening to minute 23.

Filler-word detection. The zero-cost version is a grep over your transcript:

grep -n -i -E "\b(um|uh|you know|like, like)\b" episode.srt | wc -l

That count tells you whether a de-umming pass is worth it at all (under ~40 in an hour: leave them, they’re human). MacWhisper Pro and Logic can highlight them inline. Cut the worst clusters, not every single one — surgically de-ummed audio sounds eerie.

The two-hour budget, itemized

For a 60-minute remote interview episode on one M3 MacBook Pro:

Step	Tool	Hands-on time	Machine time
Recording setup + record	Audio Hijack	65 min (the episode itself)	—
Cleanup + normalize	Logic / ffmpeg	5 min	6 min
Transcription	whisper.cpp	1 min	6 min
Edit via transcript	Logic + SRT	35 min	—
Show notes, titles, chapters, clips	Ollama + Qwen 14B	12 min	4 min
Export + upload	—	5 min	5 min

Hands-on total after the record button: just under 2 hours including the recording hour itself; the post-production share is roughly 60 minutes against the old six-hour grind. The all-free stack (QuickTime, Audacity, ffmpeg, whisper.cpp, GarageBand, Ollama) does the same job with maybe 30 extra minutes of friction. The paid stack costs about $180 once if you skip Logic.

The honest summary: AI didn’t make the podcast good — the mic, the prep, and the guest do that. What it killed is the six-to-one tax that makes people quit. Episode eleven is now a Tuesday evening, not a lost Saturday.