PDB AUDIO RESTORATION v2 — AUDIO LETTERS 21–25
Date: 2026-11-10
Input archive: Dr-Peter-Beter-HQ 03.zip
Recordings: 10; preserved decoded duration: 04:36:04.

Exact cleanup filter (FFmpeg 7.1.5):
highpass=f=60:p=2,afftdn=nr=4:nf=-50:tn=1:ad=0.85:gs=10

60 Hz two-pole high-pass; mild adaptive FFT denoising. The nr=4 setting is a
filter parameter, not evidence of a measured 4 dB improvement in every file.
No general hum notch, speech synthesis, silence trimming, time change or
dynamic compression. Some original noise and hum remain.

Full cleaned float PCM is measured using ebur128, mono (not dual-mono).
One fixed gain is applied per file:
min(-19 - integrated_LUFS, -2.5 - true_peak_dBTP, +6 dB).
The peak constraint takes priority. Actual final exported measurements are
recorded in restoration-manifest.json and QUALITY_REPORT_SK.html.

Intermediate: 32-bit floating-point WAV, mono 44.1 kHz. This prevents clipping
of cleanup transients before the measured safe gain is applied.
Export: libmp3lame, MP3 128 kbit/s CBR, mono 44.1 kHz, Xing/LAME timing metadata.
Source metadata/artwork are retained; title/comment identify the restored copy.
All MP3 outputs were decoded completely and have exactly the source sample count.
Current source and output SHA-256 values and artwork bytes were checked again
when packaging. Original source files are unchanged.

Reproduce with Python 3 and FFmpeg/FFprobe:
python3 restore_pdb_audio_v2.py --sources /path/to/Dr-Peter-Beter-HQ --manifest batch-manifest.json --output /path/to/restored --work /path/to/work --jobs 4 --stage all

Technical QC does not substitute for listening. No subjective listening review
or recovery of missing source detail is claimed.

FFmpeg reference:
https://ffmpeg.org/ffmpeg-filters.html#afftdn
https://ffmpeg.org/ffmpeg-filters.html#ebur128-1