LDM STYLE AND QUALITY: FIRST MEASURED COMPARISON
Scheduled research window: 2026-11-07T17:30:07Z
Measurement completed: 2026-11-07T17:32:29Z
Report prepared: 2026-11-07T17:34:10.654556+00:00

Purpose
Learning the requested LDM sound and quality is now the top priority of the existing music-emotion research task, especially for future transcribed-video songs. This report contains only public-safe aggregate findings. Private recording identities, source links, audio, lyrics and conditioning excerpts are intentionally excluded.

Observed, not inferred
Three authorized inventory references and one finished original generated song were measured sequentially at an owner-approved free boundary. The references were an inventory-first pilot, not a user-ranked selection or a representative sample of all LDM. Original files were unchanged.

The three reference recordings measured -12.9, -12.9 and -12.8 LUFS integrated loudness, with loudness ranges of 6.2, 6.6 and 7.4 LU. The generated comparison measured -17.7 LUFS and 9.0 LU. Thus this comparison has about a 4.8-4.9 LU integrated-loudness difference. The generated file's measured true peak was approximately -1.0 dBFS on the scanner's display. These are whole-track signal measurements, not listening ratings or proof that loudness caused dissatisfaction.

Consequence
A uniform gain of approximately 4.8 dB would exceed the generated signal's measured peak headroom. Turning it up alone is not a safe substitute for arrangement, performance or mastering work. A listening comparison can instead attenuate the louder references to a common level. Any compressor, limiter or revised mix must be evaluated separately; numerical similarity does not establish musical quality.

Method and boundaries
Existing FFmpeg 7.1.5 EBU R128 scanning with true-peak mode measured full tracks. Separate 48 kHz stereo decoding yielded sample statistics; 4096-sample Hann windows gave rough mono spectral-energy bands. Channel correlation and side-to-mid energy were recorded privately. These descriptors do not identify instruments, harmony, genre, perceived stereo width or emotional impact. No tempo, pulse, note or chord claims are made. Different track lengths and musical sections affect aggregate comparisons.

Full listening, listener preference testing and lyric-fidelity certification were not performed. ASR was not used in this measurement pass. A private, precisely documented 30-second reference-conditioning derivative was prepared for a separate authorized production pilot; that preparation is not a successful remake or proof of style transfer. The source recordings and private derivative are not distributed in this report.

Next test
Keep separate evidence for requested style, actual renderer-supported controls and heard/measured output. Compare section development, instrumental roles, vocal delivery, intelligibility, acoustics and effect transitions through full listening when available. Use one bounded production change at a time where feasible, preserve originals, and distinguish subjective preference from reproducible signal results. Do not promote the selected reference's numbers into a universal LDM target or equate greater loudness with greater quality.

Sources
FFmpeg filter documentation, ebur128 scanner and peak options:
https://ffmpeg.org/ffmpeg-filters.html#ebur128-1
EBU R 128, version 5.0, November 2023:
https://tech.ebu.ch/publications/r128
The EBU recommendation describes loudness, loudness range and maximum true peak for its programme-loudness context. Its -23 LUFS recommendation is not automatically a required artistic target for these songs.

Provenance
Measurements and source hashes are retained privately. The mathematical/psychological interpretation remains bounded: this small comparison establishes signal differences, not a universal emotional mechanism, causal diagnosis, quality pass or guaranteed future output.