calesthio/generative-media-skills/audio-mixing-mastering

Audio Mixing and Mastering Direction

Use this skill when audio must survive real playback: phone speakers, earbuds, laptops, TV soundbars, cinema-style trailers, podcasts, client review links, broadcast deliveries, and social feeds. Treat the mix as a production decision, not as a last-minute normalization step.

The job is to make the audience understand the foreground, feel the intended energy, avoid fatigue or distortion, and pass the declared delivery spec.

Start with the audio contract

Before touching levels, write a short audio contract for the project:

• Delivery context: social post, YouTube upload, podcast RSS, streaming ad, broadcast, OTT, internal review, theatrical-style trailer, music video, documentary, localization, accessibility version.
• Foreground hierarchy: dialogue/VO first, performance/music first, sound-design impact first, or a changing hierarchy by scene.
• Required deliverables: full mix, dialogue stem, music stem, effects stem, M&E, narration-only, clean captions, audio-described version, alternate language mix, stereo fold-down, 5.1/Atmos printmaster, client preview.
• Target spec: use the client/platform/broadcaster spec if supplied. If no spec is supplied, choose a conservative target and label it as a heuristic.
• Monitoring assumption: headphones-only, nearfield speakers, phone/laptop check, calibrated room, or unavailable.
• Source risk: AI voice artifacts, room noise, inconsistent clips, clipping, music licensing, generated SFX harshness, missing room tone, mono/stereo mismatch, translated VO timing, or stem bleed.

Do not master blindly to the loudest reference. Streaming and broadcast systems may normalize loudness, and over-limiting can reduce clarity while gaining little or nothing at playback.

Evidence categories

Documented facts:

• ITU-R BS.1770-5 specifies algorithms for programme loudness and true-peak signal level measurement, including K-weighting, channel weighting, gating, and true-peak guidance. Verified 2026-07-10: https://www.itu.int/dms_pubrec/itu-r/rec/bs/R-REC-BS.1770-5-202311-I!!PDF-E.pdf
• EBU R 128 version 5.0 recommends normalising programme loudness to -23.0 LUFS; where attaining target level is not practically achievable, a tolerance of +/-1.0 LU is permitted, and QC workflows may allow +/-0.2 LU for measurement error. The maximum true peak level during production linear audio should not exceed -1 dBTP. Verified 2026-07-10: https://tech.ebu.ch/files/live/sites/tech/files/shared/r/r128.pdf
• ATSC A/85:2026-07 quick reference lists -24 LKFS as the target for delivery/exchange without metadata where no prior arrangement exists, -2 dBTP maximum true peak, and a -23 to -27 LKFS range for streaming delivery services unless parties arrange otherwise. Verified 2026-07-10: https://www.atsc.org/wp-content/uploads/2026/07/A85-2026-07-Annex-M.pdf
• AES TD1008 recommends, for internet audio streaming/on-demand distribution, maximum true peak not exceeding -1 dBTP at the lossy codec input, with examples such as speech/assorted content around -18 LUFS, track-normalized music at -16 LUFS, album loudest track at -14 LUFS, interstitials at -18 LUFS, and format examples from -16 to -18 LUFS. Verified 2026-07-10: https://aes.org/wp-content/uploads/2024/01/20210924_TD1008_v3.13.pdf
• Spotify for Artists recommends targeting -14 dB integrated LUFS and keeping true peak below -1 dBTP; if a master is louder than -14 LUFS, Spotify recommends keeping true peak below -2 dBTP. Verified 2026-07-10: https://support.spotify.com/us/artists/article/loudness-normalization/
• Apple Podcasts recommends overall loudness around -16 dB LKFS with +/-1 dB tolerance and true peak not exceeding -1 dB FS, calculated according to ITU-R BS.1770-5. Verified 2026-07-10: https://podcasters.apple.com/support/893-audio-requirements
• YouTube's official upload encoding help lists recommended upload audio bitrates: mono 128 kbps, stereo 384 kbps, 5.1 512 kbps, and immersive audio at 128 kbps per channel. The same official page does not state a public loudness target. Verified 2026-07-10: https://support.google.com/youtube/answer/1722171
• Netflix branded sound mix specifications require -27 LKFS +/-2 LU dialog-gated loudness, true peaks not exceeding -2 dB True Peak, and 48 kHz/24-bit for original language mix or M&E mix masters. Verified 2026-07-10: https://partnerhelp.netflixstudios.com/hc/en-us/articles/360001794307-Netflix-Sound-Mix-Specifications-Best-Practices-v1-6
• WCAG 2.2 understanding guidance for low/no background audio says background sounds should be at least 20 dB lower than foreground speech, except brief sounds, to support users who are hard of hearing. Verified 2026-07-10: https://www.w3.org/WAI/WCAG22/Understanding/low-or-no-background-audio.html
• FFmpeg's loudnorm filter implements EBU R128 loudness normalization, supports single- and double-pass operation, can target integrated loudness, loudness range, and true peak, and upsamples to 192 kHz in dynamic mode for true-peak detection. Verified 2026-07-10: https://ffmpeg.org/ffmpeg-filters.html#loudnorm

Bundled loudness measurement tool

Use scripts/measure_loudness.py when an agent needs a repeatable local measurement before making mix or delivery decisions. It requires Python 3.11+ and an ffmpeg executable with the loudnorm filter. The script invokes FFmpeg with an argument array rather than a shell, selects the first audio stream, and emits stable JSON. It never normalizes, rewrites, or replaces the input.

Measure without assuming a target:

python scripts/measure_loudness.py final_mix.wav --pretty

Compare against an explicitly selected project target:

python scripts/measure_loudness.py final_mix.wav \
	--target-lufs -16 --lufs-tolerance 1 \
	--target-true-peak -1 --pretty

For mono files intended to play from both speakers as dual mono, request FFmpeg's dual-mono compensation explicitly:

python scripts/measure_loudness.py narration_mono.wav --dual-mono --pretty

The JSON report includes:

• integrated loudness in LUFS;
• true peak in dBTP;
• loudness range in LU;
• measurement threshold;
• analysis mode, whether dual-mono compensation was requested, and FFmpeg's reported loudnorm offset under tool metadata;
• optional target checks with actual values, deltas, tolerances, and pass/fail results;
• an explicit note that results depend on the FFmpeg build, filter, algorithm, and analysis mode.

Exit codes are:

• 0: measurement completed and every requested target check passed;
• 2: measurement completed but one or more requested target checks failed;
• 3: operational failure such as a missing input, missing FFmpeg, timeout, absent audio stream, or malformed loudnorm output.

This tool is deterministic assistance, not final mix approval. Listen to the complete deliverable, confirm the selected target is authoritative, inspect channel layout and codec behavior, and remeasure encoded outputs when delivery risk warrants it. Use a proper calibrated meter or the receiver's mandated QC system when the specification requires one.

Empirical observations:

• Small speakers often hide sub-bass and exaggerate upper-mid harshness; a mix that feels full only because of sub energy may feel thin on phones.
• Generated voices often contain transient clicks, metallic consonants, breaths in odd places, and inconsistent proximity; these problems become more obvious after compression and limiting.
• Generated music and SFX may arrive pre-limited; additional bus limiting can make them flat, gritty, or fatiguing before loudness meters show a problem.
• AI-generated ambience often loops too obviously under narration; crossfade or vary beds before mastering.

Production heuristics:

• For web/social video with speech and no formal spec, start around -16 to -14 LUFS integrated with a true-peak ceiling between -1 and -2 dBTP, then adjust for genre, platform, and client expectations. Label this as a working target, not a documented platform rule.
• For speech-first explainers, intelligibility is more important than loudness. Duck music and effects under speech before raising the whole master.
• For ads and trailers, impact comes from contrast and transient clarity, not constant maximum density. Preserve short-term dynamics around hits.
• For podcasts, listeners forgive moderate noise sooner than they forgive inconsistent dialogue level, distortion, harsh sibilance, or music masking speech.

Session prep

Make a reproducible session before processing:

1. Confirm sample rate, frame rate, timecode, duration, and target container.
2. Keep an untouched copy of every source. Never destructively overwrite original VO, music, SFX, or stems.
3. Group tracks by role: dialogue/VO, production dialogue, ADR, translated VO, music, score, stingers, SFX, ambience/room tone, UI sounds, captions/audio description references.
4. Align all audio to picture and check sync at the start, middle, and end.
5. Remove dead air only where it does not break natural timing. Preserve room tone or ambience under edits.
6. Mark problem regions: clipped words, plosives, sibilance, room noise, hum, AI artifacts, music masks, SFX spikes, silence gaps, and scene transitions.
7. Decide whether to mix from stems, flattened source, or a hybrid. If sources are flattened, note what cannot be fixed cleanly.

If receiving only a final exported video, demux audio for analysis, but keep picture reference locked. If receiving multitrack/stems, avoid mastering the stereo mix until stem balance is correct.

Balance and hierarchy

Build the mix from the foreground outward:

• Dialogue/VO: establish stable perceived loudness first. Level clip-by-clip before compression.
• Music: choose whether it supports, drives, or replaces speech. During narration, create space with level, EQ, sidechain ducking, arrangement edits, or automation.
• SFX: keep story-important effects present; trim decorative effects that compete with words.
• Ambience/room tone: fill holes and transitions, but avoid low-frequency buildup and hiss accumulation.
• Silence: use intentional silence for pacing; do not let accidental digital black expose edits.

Do not solve a balance problem only with a limiter. If the voice disappears under music, reduce or carve the music. If SFX feel too sharp, shape the SFX. If the master misses target loudness after the mix feels right, then use bus gain/limiting carefully.

Dialogue and voice treatment

For spoken-word content:

• Clip-gain first. Bring phrases into a reasonable range before inserting compressors.
• High-pass gently to remove rumble, usually below the useful fundamental of the voice. Avoid thinning baritone or warm narration.
• Cut mud/boxiness only where present; common problem areas often live around low mids, but sweep and listen rather than applying a fixed curve.
• Improve presence with small broad moves; aggressive presence boosts can make AI speech synthetic or harsh.
• Use de-essing for sibilance before final limiting. Do not let a full-band compressor overreact to esses.
• Compress in stages when needed: mild levelling, then faster peak control. Heavy single-stage compression can raise breaths/noise and flatten performance.
• Add room/reverb only when it helps the voice sit in a scene. For explainer VO, keep reverbs short and subtle.
• For avatar/lip-sync, protect consonants and timing; do not process so heavily that lip closure cues feel late or blurred.
• For multilingual dubbing, match the original scene perspective and loudness, not just the transcript timing.

If AI voice artifacts remain audible, prefer localized repairs: replace the line, regenerate the phrase, use spectral repair, trim clicks, redraw fades, or edit breaths. Broadband noise reduction on the full VO can make artifacts worse.

Music treatment

Music must be licensed, appropriate, and technically controlled: