Workflow Podcast Pipeline

Workflow — Podcast Pipeline

What: Take a raw multi-mic recording and ship a broadcast-clean, caption-complete, chapter-tagged, loudness-compliant episode to every podcast platform.

Skills used

media-whisper, media-demucs, media-denoise-ai, media-subtitle-sync, media-sox, ffmpeg-audio-filter, ffmpeg-audio-fx, ffmpeg-audio-spatial, media-ffmpeg-normalize, ffmpeg-detect, ffmpeg-metadata, ffmpeg-subtitles, ffmpeg-captions, media-batch, media-cloud-upload, media-imagemagick.

Pipeline

Step 1 — Capture + concat multi-source

Join separate mic files into aligned tracks via amix / amerge. Auto time-align against a clap or common audio marker.

Step 2 — AI denoise per stem BEFORE mixing

media-denoise-ai:

• DeepFilterNet per mic (general use, 48 kHz mono).
• RNNoise for steady-state hum.
• Resemble Enhance for speech clarity.

Step 3 — Classical audio polish

SoX or ffmpeg-audio-filter: highpass=f=80, parametric EQ boost at ~3 kHz, compand for consistent level.

Step 4 — Music bed + ducking

Sidechain compression (sidechaincompress) ducks music under voice. Order matters: key input (voice) FIRST, compressed signal (music) SECOND.

Step 5 — Transcription

media-whisper:

• large-v3 (3 GB) for best quality.
• base.en (140 MB) CPU-friendly; gap to medium (1.5 GB) is small on clean studio audio.

Use --word_timestamps True (faster-whisper) or --max-len 1 --split-on-word (whisper.cpp) for word-accurate subs.

Step 6 — Auto-sync drifted subs (optional)

media-subtitle-sync: alass first, ffsubsync as fallback.

Step 7 — Auto-chapter

• Silence-based — ffmpeg-detect silencedetect at -35 dB min-duration 5 s.
• Diarization-based — pyannote.audio or simple-diarizer (external to Whisper).

Step 8 — Loudness normalize

media-ffmpeg-normalize EBU R128 two-pass:

Step 9 — Package deliverables

• MP3 — id3v2_version 3, UTF-16 text, cover 1400–3000 px square ≤ 500 KB.
• M4A / AAC — Apple Podcasts. Chapters via @chpl atom (ffmpeg writes automatically).
• Video podcast — showwaves static waveform or pre-mix video track.

Step 10 — Batch + publish

media-batch (GNU parallel) for N-episode throughput. media-cloud-upload to host + YouTube.

Variants

• Interview podcast — pyannote diarization, SRT with speaker labels.
• Music podcast / DJ set — media-demucs stem isolation for promo clips.
• Binaural / ASMR — ffmpeg-audio-spatial sofalizer HRTF, azimuth/elevation per track.
• Video podcast with scene detection — multi-camera auto-cut at detected scenes.
• Transcript → blog post — Whisper + external LLM formatting.
• Multi-language release — transcribe original, translate externally, mux multiple sub tracks with language metadata.
• AI-generated intro — TTS branded intro + Riffusion bed + mix. See workflow-ai-generation.

Gotchas

• Whisper wants 16 kHz mono — resamples internally, but explicit -ar 16000 -ac 1 is cleaner.
• Demucs wants 44.1 or 48 kHz STEREO. Mono input → degraded stems.
• DeepFilterNet: 16 kHz or 48 kHz MONO only. Mix inputs at 48 kHz mono.
• RNNoise: 48 kHz mono 16-bit, strict. Resample first or it fails silently.
• loudnorm -target -14 is Spotify Master, NOT podcasts. Podcasts want −16 (Apple) or −19 (ACX).
• Single-pass loudnorm uses guardrails — two-pass measures then applies. For compliance, use media-ffmpeg-normalize (two-pass).
• loudnorm applies a limiter — over-loudnormed audio sounds compressed. Tune lra to preserve dynamics.
• Sidechain compression order: key input FIRST, compressed signal SECOND. Wrong order = wrong ducking.
• Whisper hallucinates on silence. silenceremove leading/trailing before transcription.
• Whisper timestamps end-of-chunk by default. Word-level requires --word_timestamps.
• Whisper --language auto is fragile on accents. Specify explicitly.
• MP3 ID3v2.4 vs v2.3: Apple Podcasts requires v2.3 + UTF-16 text. -id3v2_version 3.
• Cover art: 1400–3000 px square, < 500 KB JPEG/PNG. Apple rejects larger.
• MP3 chapter metadata uses CTOC + CHAP frames. ffmpeg writes from -metadata chapters but older players ignore.
• M4A chapter metadata uses @chpl atoms — ffmpeg writes automatically.
• RSS enclosures need exact file size + MIME type. Post-publish file change = hash mismatch → listeners re-download.
• Serialize loudness normalization if sharing GPU with other jobs — two parallel loudnorm = memory contention.
• alass fails on already-synced audio (returns input unchanged). --force won't help.
• ffsubsync pattern-matches audio to existing subtitle anchors. Naive use on very-drifted subs = drift amplification.
• showwaves locks to one audio stream. For multi-stream, use showwavespic (static) or pre-mix.

Example — 60-minute interview → deliverables

Multi-mic input → DeepFilterNet per mic → amix → EQ + compand → music bed with sidechain duck → Whisper large-v3 word-timestamps → alass subtitle sync → silencedetect auto-chapter → loudnorm −16 LUFS two-pass → MP3 + M4A with chapter atoms + cover art → upload.

Related

• workflow-audio-production — deeper audio-chain recipes.
• workflow-ai-generation — AI intro / voice clone / thumbnails.
• workflow-analysis-quality — loudness compliance verification.