mix-qc-and-delivery

The objective QC gate and delivery specs for a spoken-word mix (documentary, radio feature, podcast): dialogue-anchored loudness, true peak at 4x oversampling, LRA, dead air, clicks at cue edges, mask

SKILL.md

GitHub

mix-qc-and-delivery

The mix is checked before any picture is rendered. A late mix failure after a picture render cost hours in practice; the gate exists so that cannot happen again. OpenDubStage tool: opendubstage qc (opendubstage.qc.mixqc: qc(mix, cues, stems, thresholds), Thresholds).

opendubstage qc MIX.wav [--cues SHEET.json] [--stem dialogue=DX.wav --stem background=BG.wav] [--thresholds T.json] [--json OUT.json] [--md OUT.md]

Exit 1 when a blocking check fails. Without a cue sheet only the programme-wide checks run (loudness, LRA, true peak, dead air). Cue-sheet schema and the full threshold table with sources: references/thresholds.md.

1. Delivery specs

Spec Integrated True peak Notes
PRSS (public radio distribution) −24 LUFS ± 2 −2 dBTP (sample −3) Current 2014; Radio World
ATSC A/85 −24 LKFS ± 2, dialogue-anchored −2 dBTP
Netflix v1.5 dialogue −27 LKFS ± 2, dialog-gated −2 dBTP LRA programme 4–18 LU, dialogue ≤ 7 LU (guidance)
EBU R128 −23 LUFS ± 1 −1 dBTP LRA descriptive; talk ≈ 5 LU (Auphonic)
Apple Podcasts −16 LKFS ± 1 −1 dBTP
AES TD1004 streaming −16…−20 LUFS ≤ −1 dBTP lossy codecs overshoot
Home-viewing feature (recommended) dialogue −24 LUFS dialogue-gated ± 1; programme free (expect −22 ± 3) ≤ −2 dBTP after the codec; ≤ −2.3 before AAC LRA 8–18 LU reported, not clamped; dialogue-only LRA ≤ 7

Anchor on dialogue, not on the programme: measure dialogue-gated loudness on the dialogue + live-speech stems, apply one linear gain, then a limiter as the only non-linear stage. Never loudnorm in dynamic mode on a long mix: ffmpeg falls back to dynamic (3 s window compression) whenever the measured LRA exceeds the target LRA; check the first pass's normalization_type.

2. True peak

A sample-peak limiter misses inter-sample overs. Measured in practice: a mix limited at −2.5 dBFS sample read −1.3 dBTP. Rules:

  • Limit at 4× oversampling (ITU-R BS.1770-4) or set the sample ceiling 0.6 dB below the target: alimiter=limit=0.74 (−2.6 dBFS) to land ≤ −2.3 dBTP before AAC, ≤ −2 after.
  • Verify with ebur128=peak=true or opendubstage qc (true_peak_max −2.0), never with the limiter's own meter.
  • Leave 0.3–0.5 dB for the codec; re-measure the encoded file.

3. The checklist

# Check Threshold Gate Source
Q1 dead air no span < −60 dBFS sample peak for > 3 s (strict: −50 dBFS / 2 s), head and tail 5 s exempt, designed silences whitelisted blocking silencedetect; Torcoli (ambience ≥ 15 LU under speech)
Q2 gaps are air momentary (400 ms) ≥ −46 LUFS in every speech gap; ambience in the clear −38…−42 LUFS blocking Torcoli; Tech 3341
Q3 dialogue anchor dialogue-gated −24 ± 1 LUFS blocking PRSS; A/85
Q4 programme integrated −27…−16 (report); LRA 4–18 LU; dialogue-only LRA ≤ 7 (narrated segments ≤ 10, live ≤ 15 advisory) advisory Netflix; Auphonic; Tech 3342
Q5 true peak ≤ −2.0 dBTP at 4× oversampling (−2.3 before AAC) blocking PRSS; Netflix; BS.1770-4
Q6 every voice at the same level short-term (3 s) per voice class within ±2 LU of −24; per-cue sentence std ≤ 1.5 LU advisory Byers
Q7 tonal balance 1–4 kHz presence share per voice class within ±4 dB of the class mean; LTAS per class within ±3 dB of the reference 200 Hz–6.3 kHz advisory RX EQ Match practice; measured in practice
Q8 rumble < 100 Hz ≥ 25 dB below the 1 kHz band on dialogue advisory measured in practice
Q9 masking music + ambience ≤ dialogue −10 dB in 1–4 kHz over any ≥ 3 s stretch of speech (strict: −14 dB in 1.2–4.8 kHz, 99th percentile); broadband music ≤ −10 LU, ambience ≤ −15 LU blocking (default) Torcoli; Shirley & Kendrick
Q10 no vocals under speech no under cue overlapping speech > 1 s whose section has vocals blocking (plan) NPR practice
Q11 music entries downbeat / section / song start ≤ 60 ms; fade_in ≥ 0.5 mid-song; none within 2 s before a first word blocking (plan) Menjivar
Q12 music exits / posts every under cue ends on a downbeat or section end with ≥ 1 s out; posts 2–12 s; music-only > 15 s flagged advisory Menjivar
Q13 music density per 10 min: 30–70 % music; no cue > 150 s under speech; no dry stretch > 90 s without bed advisory Menjivar
Q14 crossfade at every boundary ±3 s: a non-dialogue stem non-zero on both sides ≥ 1.5 s; music adjacency gap < 0.1 or ≥ 3 s or overlap ≥ 1.5 s blocking check_transitions
Q15 edit points every dialogue cue fade_in ≥ 0.05, fade_out ≥ 0.25, last_word_end + 0.35 ≤ cue_end blocking Audition 5 ms rule; measured in practice
Q16 clicks sample step > 8× the local difference RMS within ±5 ms of a cue edge, above −60 dBFS; adeclick save-mode < 1 click/min outside repaired events blocking engineering practice
Q17 speech heard correctly transcription check: head and tail ratio ≥ 0.5 for every narration cue; key live phrases ≥ 0.75 blocking measured in practice
Q18 event residue energy inside each repaired event window within 3 dB of the floor advisory
Q19 restoration safety DNSMOS SIG not lower than source; speaker-embedding cosine ≥ 0.9; texture floor Δ ≤ 6 dB; WER on secondary speakers unchanged blocking for any neural pass Smule 2025 metrics; measured in practice
Q20 half-volume test mix attenuated 12 dB plus a −50 dBFS pink floor: Whisper WER on 10 random 30 s windows ≤ 1.3× the WER at full level advisory Byers: "turn it down halfway, can you understand every word?"
Q21 level jumps > 5 LU between adjacent voices' momentary max within 1.5 s of a voice-to-voice edit; scene starts exempt advisory JND ~1 LU; A/85 ±2 LU
Q22 hard stops music / ambience / effects fade-out < 1.5 s while still sounding (programme end exempt) advisory Byers
Q23 mix length mix = timeline ± 0.05 s; picture later = audio ± 0.05 s blocking
Q24 anti-fatigue every 10-min window has ≥ 2 of {music enters/exits, bed change, live sound, ≥ 90 s music-free} advisory Kwok, Get Back
Q25 separated speech keeps the room residual stem ≥ −20 dB vs separated speech in every separated span advisory Get Back

Blocking by default in Thresholds: loudness, dead_air, click, masking. Tune per programme; every threshold carries its source in the dataclass.

4. Two rules learned the hard way

Dead air comes from beds shorter than their scenes. Before mixing, check every bed's source length against the span it must cover (loop it, or crossfade a second bed); then run the detector. Designed silences are whitelisted explicitly (a list of (at, dur)); the dead-air check reads the list; nothing else passes.

Only hard edits can click. A cue edge with a fade ≥ 50 ms cannot produce a step discontinuity; a transient found at a faded edge belongs to other content (a consonant, a spot effect, the music) and is not a click. Report clicks only on edges whose fade is < 50 ms, and report the fade length with the finding so the reader can tell the two apart.

5. The transcription check (every cut)

Transcribe each narration cue out of the finished mix (not the source) and compare its first and last six words with the word-timed transcript. Procedure in references/transcription-check.md. Essentials:

  • normalise both sides: lower-case, strip punctuation, drop fillers (um uh er ah hmm mm), drop false starts (tokens ending in -, e.g. "bea-"), collapse stutter repeats (Whisper smooths these; the transcript keeps them; neither is a clipped word);
  • overlapping cross-talk is optional: Whisper drops the other speaker, so build two expectations, all speakers and the main speaker of the cue (most common label), and pass if either matches (diarisation labels are unreliable);
  • difflib.SequenceMatcher ratio on want[:6] vs got[:8] and want[-6:] vs got[-8:]; pass when head ≥ 0.5 and tail ≥ 0.5;
  • check every cue: a 12-cut sample missed "and--" endings a model reviewer heard;
  • for live scenes, check that a handful of key phrases are heard (fuzzy ≥ 0.75).

6. The humans

  • Multi-system: a 90 s excerpt bounced through a phone-speaker EQ (highpass=300, lowpass=6000), a laptop and the living-room TV, at normal volume (Tavella checked The War on six to eight systems down to an 11-inch mono TV).
  • Headphones: three 60 s excerpts (opening, a bed-heavy passage, the ending) on open-back headphones before lock; −40 LUFS beds and spot effects that vanish on a TV are plainly audible there (Mangini, 32 Sounds).
  • Half volume: Byers' test, by ear as well as by Q20.
  • The witnesses sign off: the people who were there decide, on their own system.

7. What the automated gate does not hear

Measured in practice against a model reviewer's 139 relevant notes: the checks caught 38 (dead air 21/31, masking 9/25, level 8/29, abrupt edges 5/47, tone 6/49). The misses were content edits (mid-sentence cuts: covered by the transcription check once it runs on every cue), local rumble (the tonal check compares class averages, not moments), sung vocals under speech (energy-based masking is not vocal-aware) and music chopped mid-phrase with a ≥ 1.5 s fade still present. Keep the transcription check, the model listener (ai-review-of-audio) and the human listens in the gate; the numbers alone pass mixes a professional would not.

Known blind spot: a cue wrongly tagged as non-speech (e.g. a live-music bypass applied to a speech) drops out of the duck sidechain and out of every dialogue check. Assert the count of bypassed cues in the build, and that known speech sources are not among them.