Run a healing app long enough and "sound" climbs into the surprising top ranks of reasons a star gets docked. "I used it before bed and it was suddenly loud," "I thought it was silent, then it played a few seconds in" — these quietly erode ratings more than functional bugs do.
Audio assets don't announce themselves with a compile error the way code does. Verifying every one by ear becomes unrealistic as the track count grows. With the 6/26 update, Antigravity can render audio files as waveforms inside the IDE, opening a path to offload part of this "ear work" to eyes and scripts. As someone who has shipped sound in indie apps, here is a design that systematizes pre-release audio checks as a verification loop.
Where audio assets actually cause accidents
First, let's make concrete which defects break the reader's experience. A vague "quality check" hands an agent a vague result.
Across my operations, the ones most likely to draw complaints and low ratings were four: loudness varying between tracks, peaks clipping into distortion, unnatural silence or an abrupt onset at the start, and loop material whose tail and head don't join smoothly.
Most of these are visible in the waveform. Put differently, if you reduce them to things a waveform can show, you can give the agent a footing for judgment.
Kill the machine-decidable parts first
Before sending anything to ears or an agent, automate what a number can decide. Skip this and reviews drown in subjective back-and-forth. I always run a light gate built on ffprobe and ffmpeg's loudnorm analysis before publishing.
#!/usr/bin/env bash
# audio-gate.sh — machine-judge audio assets before release
# usage: ./audio-gate.sh assets/sounds/*.wav
set -euo pipefail
# distribution targets (tune to your app)
TARGET_LUFS=-16 # a mobile playback ballpark
MAX_TP=-1.0 # true-peak ceiling dBTP
MIN_SEC=1.0 # reject material that's too short
MAX_HEAD_SILENCE=0.3 # allowed head-silence seconds
for f in "$@"; do
dur=$(ffprobe -v error -show_entries format=duration -of csv=p=0 "$f")
stats=$(ffmpeg -i "$f" -af loudnorm=I=$TARGET_LUFS:TP=$MAX_TP:print_format=json -f null - 2>&1 \
| sed -n '/{/,/}/p')
lufs=$(echo "$stats" | python3 -c "import sys,json;print(json.load(sys.stdin)['input_i'])")
tp=$(echo "$stats" | python3 -c "import sys,json;print(json.load(sys.stdin)['input_tp'])")
head_sil=$(ffmpeg -i "$f" -af silencedetect=noise=-50dB:d=0.1 -f null - 2>&1 \
| grep -m1 "silence_end" | grep -oE "[0-9.]+" | head -1 || echo 0)
status="OK"
awk "BEGIN{exit !($dur < $MIN_SEC)}" && status="TOO_SHORT"
awk "BEGIN{exit !($tp > $MAX_TP)}" && status="CLIPPING"
awk "BEGIN{exit !($head_sil > $MAX_HEAD_SILENCE)}" && status="HEAD_SILENCE"
printf "%-28s dur=%.1fs LUFS=%s TP=%s headSil=%ss -> %s\n" \
"$(basename "$f")" "$dur" "$lufs" "$tp" "$head_sil" "$status"
doneJust running this surfaces, in one list, problems that are hard to catch without A/B listening: "one track is 6 dB louder," "the tail is clipping." When I swapped an AdMob reward sound, this stage caught a true-peak overage and I fixed it before release.