nvrl
15853e61ac
3.6 KiB
Devath reference measurements
Analyzed the seven user-supplied FLACs in this workspace using their actual audio samples. This is signal analysis of the rendered mixes, not a listening-based transcription of their melodies or an account of the artist's intent.
Observable contrasts
Approximate reference points from the one-second summaries:
| Recording | Measured pattern | Useful compositional question |
|---|---|---|
| Henka | A marked reduction in level/attacks around 2:03–2:35, followed by a strong return; another reduction near 3:57. | What can be withheld so the recognizable idea matters when it returns? |
| Phantasy | Lower, interrupted activity around 1:27–2:23, contrasting with the stronger regions before and after. | Can a fragment carry continuity across a less active passage? |
| Blumengarten | An interrupted middle region around 2:02–3:02, then renewed sustained activity; marked reduction near 4:19. | Can the same identity support different densities and articulations? |
| quiet wish | A broad reduction around 1:53–2:32 before activity returns. | Does a long breath make the return more effective than adding notes? |
| Light Aspect | A smaller reduction around 1:33–1:49 and a later reduction around 3:04. | Can a shorter contrast work without rebuilding the entire phrase? |
| Petals | More consistent relative activity through much of the first four minutes, with short dips and a lower closing region after roughly 4:11. | Can development happen with small changes while the groove stays continuous? |
| VERDICT. remix | Stronger sustained regions roughly 1:01–1:48 and 3:02–3:48, separated by a less active region. | What should remain recognizable across contrasting presentations? |
These times are prompts for human listening, not asserted verse/drop boundaries. RMS is affected by arrangement and mastering. Spectral attacks include drums and sound-design transients, so neither curve measures melodic note density.
The design inference is to support both subtraction/return and continuous-groove writing, rather than force every generated phrase through the same climb and resolution. The implementation includes optional fixed groove, fragments, note subtraction, different articulations, register return, and section composition. This does not establish that Devath used those exact note-level operations. No source notes or audio were copied into presets or used to train a model.
Reproduce
python scripts/analyze-references.py references/devath --output docs/rivet/research/measurements
Requires ffmpeg, NumPy, and Matplotlib. The script decodes mono audio at 11,025
Hz and uses a 2,048-sample Hann window with a 256-sample hop. It summarizes RMS
and positive log-spectral flux per second, normalizing each curve to its own
95th percentile for visibility. Candidate texture changes compare standardized
low/mid/high-band energy over eight-second neighborhoods, with separated peaks.
Raw summaries include track durations and
within-track 80th/20th percentile ratios. Automatic tempo and chord estimates are
intentionally omitted: coarse onset periodicity and spectral peaks are not
reliable transcriptions of these full mixes.
A detailed melodic study still needs a listening pass with annotated motifs, harmonic rhythm, bass/lead relationships, and returns, ideally checked against stems or MIDI. These measurements alone cannot identify exact chord progressions, voice leading, emotional meaning, or which sound carries the melody.