Workflow — Audio Production

Step 2 — Control surfaces

• MIDI — media-midi: list-ports, monitor JSON, send-note, map to OBS/OSC/anything.
• OSC — media-osc: send/listen.

Step 3 — DSP chain

• SoX / FFmpeg filters — highpass=f=80, equalizer=f=3000:t=q:w=1:g=3, acompressor, alimiter, loudnorm.
• LADSPA / LV2 plugins — via ffmpeg's ladspa / lv2 filters.

Step 4 — Spatial audio

ffmpeg-audio-spatial:

• Binaural HRTF — sofalizer with a SOFA file, per-track azimuth / elevation.
• Surround 5.1 upmix / downmix — pan filter with channel matrix.

Step 5 — AI processing

media-denoise-ai (DeepFilterNet / RNNoise / Resemble Enhance), media-demucs (stem separation), media-tts-ai (Kokoro / OpenVoice / Piper), media-musicgen (Riffusion / YuE), media-whisper (transcription).

Step 6 — Loudness certification

media-ffmpeg-normalize two-pass EBU R128:

Variants

• Live DAW recording — start JACK, connect DAW I/O, MIDI controller, OSC tablet.
• Podcast cleanup — see workflow-podcast-pipeline.
• Binaural audiodrama — split stereo, per-track HRTF positioning, sum.
• DAW → OBS system-audio routing — aggregate device, DAW out → BlackHole, OBS source → aggregate.
• Control-surface-driven automation — Stream Deck MIDI → OBS scene + DMX cue + OSC transport in parallel.

Gotchas

• BlackHole / Loopback appear as BOTH input AND output on macOS. Route the app to OUTPUT; pick input in DAW/OBS.
• PipeWire aliases default to the current Pulse sink. Script with the real sink name from pw-cli ls Node.
• JACK buffer = latency frames ÷ sample rate (128 @ 48 kHz = 2.67 ms). Too low = xruns; too high = sluggish.
• WASAPI loopback captures the render endpoint, not the source app. Use a virtual cable for per-app capture.
• macOS aggregate devices clock to the FIRST device in the list. Others drift. Pick the clock master first.
• Windows MME is legacy shared; WASAPI is exclusive/professional. For low latency, use WASAPI.
• MIDI 1.0 = 3 bytes; MIDI 2.0 UMP = 4–16 bytes. Wrong protocol = silent ignore.
• MIDI channels: 0–15 on the wire, 1–16 in UI. Off-by-one is the #1 bug.
• MIDI running status compresses repeated types — old hardware uses it; modern mostly doesn't.
• OSC bundles contain timed sub-messages. Parse type:bundle → .elements branch.
• OSC address patterns support wildcards (/track/*/volume). Server-side wildcards fire multiple handlers.
• OSC TCP 1.1 vs UDP 1.0. Pick one per endpoint; they're not interchangeable.
• loudnorm is NOT a limiter. It normalizes. If peaks exceed target, it applies downward gain, not transient shaving. Add explicit alimiter.
• aresample filter_size trades quality for speed. 512 = audiophile, 128 = streaming.
• atempo range is 0.5–2.0 per instance. Chain for extremes.
• asetrate changes pitch AND tempo together. Combine with atempo for pitch-preserving speed.
• sidechaincompress — key input FIRST, signal SECOND. Wrong order = wrong ducking.
• DeepFilterNet: 48 kHz MONO or 16 kHz MONO. Resample before.
• RNNoise: 48 kHz mono 16-bit, hardcoded. Anything else fails silently.
• Demucs: 44.1 / 48 kHz STEREO. Mono → degraded stems.
• Whisper: 16 kHz mono best. Resamples internally; explicit is cleaner.
• Voice cloning needs a CLEAN reference. DeepFilterNet the reference sample first.
• Kokoro TTS outputs 24 kHz. Resample to 48 kHz before mixing.
• Single-pass loudnorm uses guardrails (±1 LUFS). Two-pass measures then applies exactly. Use media-ffmpeg-normalize for certification.
• True peak measures inter-sample peaks, 4× oversampled. Standard sample peak misses these.
• LUFS = LKFS = same thing (different standards' names for integrated loudness).

Example — Podcast-style live monitor with MIDI transport

jackd start → connect DAW ↔ system audio → media-midi monitor Stream Deck → MIDI CC maps to DAW transport (play/stop/record) AND to OBS scene switch in parallel.

Related

• workflow-podcast-pipeline — applied podcast recipe using this stack.
• workflow-live-production — MIDI/OSC surfaces driving OBS + DMX + PTZ.
• workflow-analysis-quality — loudness compliance verification.