800 lines went into the engine, and QA said two things: "some are hard to hear, some are blown out." That is a loudness problem, not a performance problem.

https://apps.microsoft.com/detail/9pn6cswk4jpz?cid=DevShareMCLPCS&hl=zh-CN&gl=CN
If your game VO sounds uneven, nine times out of ten it is not the recording — it is that no one aligned every file to the same loudness target. Hearing is not linear: two lines peaking at -30 dBFS, one averaging -24 LUFS and the other -17 LUFS, will read as "clear and emotional" versus "muffled". Then the explosion hits, the music swells, and the -17 one disappears under the mix.
Three numbers decide this: average loudness (LUFS), true peak (dBTP), and how consistent they are across the batch.
Three familiar failures
Peak-only thinking. Bring every file's peak to -1 dBFS and call it done. Now the quiet, intimate lines are loud and the shouted ones are quiet — peaks describe one instant, not perceived loudness.
Brute-force compression. Compression shapes tone; it does not align loudness. Push it and sibilance and breath come up with everything else, and the performance flattens.
Hand-tuning 800 files. The first twenty get real attention. By file 200 you are guessing. By 800 you just want to ship. That is exactly where consistency dies.
The order that works: normalize → limit → unify
Loudness normalization to the target — game VO is usually -23 LUFS, the EBU R128 broadcast figure.
True-peak limiting at -1 dBTP, with 4x oversampled detection, leaving headroom for encoding.
Unify the spec — sample rate, bit depth, channel count — so the engine never sees a 44.1 kHz mono file next to a 48 kHz stereo one.
Naming and folders that match your Wwise or FMOD hierarchy, so import needs no cleanup.
Denoise and de-ess belong before normalization. Do them after and the normalizer simply lifts the tail you just removed.
Doing it in LoudPipe
That sequence is a single preset: game_vo_lufs (-23 LUFS / -1 dBTP). Drop the vo/ folder on the window and press start.
Measured within 0.1 LU — not "close enough", every file lands in the same narrow band.
Failure isolation — one corrupt file among 800 does not stop the other 799; "Retry failed" fixes just that one.
Optional cleanup — the game_vo_clean preset adds spectral denoise and dynamic 5–8 kHz de-essing, skipping clean material and material with no sibilance instead of smearing it.
Reports come with the batch — input/output LUFS, true peak, clipping, processing time per file, in CSV / JSON / HTML. When QA asks why one line sounds small, you hand them numbers.
Concurrency and retries — -j 4 --retries 2, configurable in the GUI too.
One detail people skip
Do not stop at the green "success" column. Listen to three files: the quietest, the loudest, the most sibilant. Then look at the band-energy view — sub, low, mid, presence, sibilance, air plus peak frequency. Whether that sibilance is genuinely harsh stops being an opinion.
Loudness is invisible when it is right and audible to everyone when it is wrong. Control the three numbers, and 800 files is just one batch.
LoudPipe is on the Microsoft Store: batch loudness, true peak, denoise and de-ess, text to speech and audiobook delivery in one window.
https://apps.microsoft.com/detail/9pn6cswk4jpz?cid=DevShareMCLPCS&hl=zh-CN&gl=CN

Top comments (0)