How one broadcast team cut QA review time from hours to minutes

Before automating any of it, a full episode delivery went through a manual listening pass against the mix reference — a real person, headphones on, for the full runtime of every episode, every delivery format.

Before: a full listening pass, every time

For a weekly series with three delivery profiles — broadcast, streaming, and a loudness-adjusted regional variant — that meant roughly three full-runtime listens per episode, most weeks confirming that nothing had changed. At an average 42-minute runtime across three profiles, that’s over two hours of listening time per episode, nearly all of it spent reconfirming a delivery that turned out to be fine.

What automation actually replaced

The pipeline now runs wavelens analyze against the mix reference for every delivery profile automatically, with wavsummary rolling the results into a single gate before a delivery is marked ready:

$ for profile in broadcast streaming regional; do
    wavelens analyze \
      --ref mix_reference.wav \
      --dut delivery_${profile}.wav \
      --out reports/${profile}.html --json
  done
$ wavsummary ./reports --json

A clean run — all six checks passing across all three profiles — ships without a listening pass at all. That’s the majority of deliveries now, since most delivery-profile transcodes genuinely don’t change anything Level Fidelity, Spectral Fidelity, Distortion, or Loudness would catch.

What actually moved the number

The workload change wasn’t replacing human listening — it was making it selective. A failed check routes the episode to a human reviewer with the report already open to the specific section and check that failed, instead of a blind full-runtime listen across all three profiles. Reviewing a report with a flagged discrepancy — reading the failed check, jumping to the timestamp it names, confirming by ear — takes minutes. A full blind listening pass across three delivery profiles took the better part of two hours.

That’s the actual shift: from “listen to everything, every time” to “listen to the part a report already pointed at, when there’s a reason to.”

What still gets a human ear

The team still runs a full manual listening pass on a sampling basis — a handful of episodes a month, regardless of whether the automated checks passed — specifically to catch what six checks structurally can’t: a mix decision that’s technically unchanged but creatively wrong, or a class of problem nobody’s written a check for yet. Automation narrowed where human attention goes; it didn’t remove the need for it.