The Sneakiest AI Video Failure: Our Own Fade-Out Setting Strangled the Narration (QA Checklist Included)

This video’s QA report was all green: resolution passed, frame rate passed, subtitles passed, “no clipping” passed, “audio present” passed. But the narration was silent — all 17 narration segments were barely audible from start to finish, because one fade-out parameter in the assembly script had its start point wrong by the length of a segment. This is the true story of Veiled Lore Atlas’s Doll Island V1, and the reason per-segment volume verification is an iron rule here now.

ENVIRONMENT

Veiled Lore Atlas, Doll Island V1 (1080×1920 vertical, 20 shots, 110.9 seconds total, 17 narration segments)

TOOLS

ffmpeg audio assembly (narration + brown-noise bed + wooden-clapper effects), automated QA checks

DATE

2026-10-12 (the all-pass QA came first; the audio repair record is from the same day)

KEY FIGURES

before repair, a narration segment (4–14 s) measured mean −51.8 dB; after repair, five sampled segments measured mean −16.8 to −18.3 dB

The scene of the failure: QA all green, narration silent

Doll Island V1’s automated QA report (2026-10-12) read like this: 1080×1920 ✓, 30 fps ✓, all 20 sampled frames unique ✓, 50 subtitle lines ✓, no Traditional-character violations ✓, no clipping (max −2.8 dB) ✓, audio present (mean −31.7 dB) ✓ — ALL PASS.

The problem: that −2.8 dB peak came from the wooden-clapper effects and the ambience bed, not from narration. The −31.7 dB overall average was mostly the noise bed contributing. The narration track’s actual state was measured in the follow-up investigation: in the narration mix, the 4–14 second segment had a mean of just −51.8 dB and a max of −18.7 dB — by any human standard, silence. Every green light on the report was genuine; none of them was measuring whether the narration existed.

The root cause: a fade-out that starts at second zero

When the assembly script extracted each narration segment, it applied a closing fade: afade=t=out:st=0:d=0.15. The intent was “fade out over the last 0.15 seconds of the segment”. But st=0 means the fade starts at second 0 of the segment — so every narration segment began decaying toward silence the moment it started, reached it within 0.15 seconds, and the sentence was strangled. All 17 segments, every one.

During the repair we ran a controlled measurement to quantify what that one parameter does:

Same audio file, same narration segment — changing only the fade’s start point moved the average level by about 48 dB. This class of parameter error is the sneakiest kind: it throws no error, crashes nothing, and produces an output of exactly the right duration that looks fine in every respect. Only listening — and measuring segment by segment — catches it.

The repair: rebuild the audio only, touch no video

The repair was deliberately scoped small: rebuild only the audio track (17 narration segments re-aligned, loudness normalization, the noise bed and clapper effects carried over from the original), copy the video stream through untouched, total duration unchanged at 110.9 seconds. Verification no longer looked at overall figures; it measured five narration segments individually:

All five passed the new thresholds (mean > −25 dB, max > −10 dB). The record even keeps an honest footnote: the 20–30 s segment’s max of −2.2 dB sits 0.3 dB above the −2.5 dB target — inaudible, no clipping risk, ruled not worth a re-run. A QA record that writes down even that is a record you can trust.

The checklist: sign off narration audio segment by segment

This checklist has been fixed procedure ever since. The next production (Minxiong Haunted House EP02) wrote “compute the fade-out start from the measured narration length” directly into its assembly script comments, aimed at this exact pit:

1. Measure per segment, never overall. Run volumedetect on every narration segment: mean > −25 dB, max > −10 dB. Overall peak and mean are context, never proof that narration exists.
2. Sample across the whole runtime. At minimum five segments spanning head, middle and tail (we sampled 4–14 / 20–30 / 40–50 / 55–65 / 98–108 s) — never just the first sentence.
3. Audit every fade’s start point. Every afade start must be computed as segment length minus fade duration; a within-segment fade with st=0 is forbidden. Print the parameters and read them before assembling.
4. Attribute the peaks. Know who contributes the loudest moment: if the peak belongs to an effect (clapper, rain), it says nothing about narration.
5. Measure the fix against the bug. Before and after any parameter change, measure the same segment and write both numbers into the record. “Sounds fixed” is not verification.
6. Log the small deviations too. A 0.3 dB judgment call gets written down, so the next person verifying knows it exists.

Failures and limits: what this checklist can’t catch

House rule — this section is never skipped:

• The checklist was bought after the fact. V1’s silent narration was only discovered after an all-green QA — the check suite simply had no item for “narration doesn’t exist” as a failure mode. Your checks define what you can see; invisible failures don’t announce themselves.
• The thresholds are empirical. Mean > −25 dB and max > −10 dB are practical thresholds for a narration track, not an industry standard; a mix where a music bed fights the voice needs separate judgment.
• Volume says nothing about content. This check proves the narration is present and loud enough. Whether the words are right is a different verification — the character-by-character one in article 3. The two checks complement each other; neither substitutes.
• V1 has been superseded. Doll Island was later remade as V2 under the storyboard-first workflow and published; V1’s remaining value is this lesson record, and every figure in this article comes from V1’s QA and repair files.

Sources and further reading

• Doll Island V1 QA report and audio repair record, qa-report.txt (2026-10-12): the all-PASS report, the afade root cause, the controlled before/after measurements and the five-segment verification figures are from here.
• Doll Island V2 assembly scripts (step2_assemble.py / step4_audio.py): the “compute fade start from segment duration to avoid the V1 afade bug” comment and implementation are from the V2 production files.
• Minxiong Haunted House EP02 assembly report (2026-10-14): the fade-out start computed from the ffprobe-measured narration length (160.836 s) — the same pit-avoidance pattern carried forward.
• Article 3 on this site, “ElevenLabs Chinese Narration, Tested”: the character-by-character verification (are the words right) complements this article’s volume verification (is the voice there at all).

FAQ