A research diary

The Journal.

What was tried, what was measured, and what it taught us. Findings are reported as they land, including the disappointing ones; that is the point of the instrument.

30 Aug 2026

The loops were in the decoder

Op. 10's raw samples looped as badly as op. 9's, which briefly read as a failed study. A sampling sweep told a different story: the model knew more music than its decoding let out. Sharpened by training, unpenalised sampling kept collapsing it into its favourite figures. The full grid, two styles by three temperatures by three repetition penalties, 500 tokens per run, fixed seed. rep4 is the fraction of repeated four note figures; the real corpus averages 0.54 with about 12 unique pitches per melody.

Style	Temp	Penalty	Notes	Pitches	Durations	rep4
pop	0.9	1.0	166	9	5	0.85
pop	0.9	1.2	241	15	5	0.54
pop	0.9	1.4	229	10	11	0.34
pop	1.1	1.0	249	3	3	0.94
pop	1.1	1.2	230	12	21	0.22
pop	1.1	1.4	203	16	18	0.19
pop	1.3	1.0	189	17	26	0.39
pop	1.3	1.2	189	32	20	0.06
pop	1.3	1.4	188	22	20	0.13
jazz	0.9	1.0	167	6	3	0.90
jazz	0.9	1.2	214	14	13	0.27
jazz	0.9	1.4	198	18	14	0.18
jazz	1.1	1.0	201	10	12	0.76
jazz	1.1	1.2	203	30	6	0.07
jazz	1.1	1.4	207	20	16	0.09
jazz	1.3	1.0	185	22	16	0.26
jazz	1.3	1.2	198	26	16	0.04
jazz	1.3	1.4	211	25	13	0.01

Serving settled on penalty 1.2 near temperature 0.95: repetition lands at the corpus statistic without drifting into the incoherence the hottest settings show. The full listening batch regenerated at those settings, three melodies per style, 600 tokens each:

Style	Notes	Range	Pitches	Durations	rep4
classical	240	56 to 101	26	9	0.11
rock	251	54 to 94	20	22	0.16
dance	262	59 to 98	24	11	0.19
pop	277	53 to 90	17	11	0.34
rnb	252	62 to 96	17	13	0.34
folk	269	40 to 93	22	11	0.34
jazz	281	38 to 83	18	7	0.39
standards	290	41 to 83	16	15	0.61
real corpus			12	38	0.54

Every style now exceeds the corpus in pitch variety, and the styles separate: classical roams furthest, standards keeps the tightest song form repetition. Op. 10 serves on the desk with the tuned decoder, and a study that looked disappointing became the best finding of the week for the price of a five minute experiment.

30 Aug 2026

Op. 10 measures the scaling slope, and it is shallow

The tenth study trained on 30,007 melodies, 1.6 times the ninth corpus, on a rented A100. Validation loss reached 1.30 against op. 9's 1.33 before the same overfitting pattern returned. The finding: at 34M parameters, raw data volume alone buys little. The next study attacks quality instead, filtering the corpus by the same repetition and variety metrics we judge output with, and training with heavier regularisation. A disappointing number, and a useful one.

30 Aug 2026

Op. 9 goes public, honestly labeled

The ninth study is live on the composer's desk as a research preview: the first Melograph model with working style conditioning. Choose a genre and the model is steered by a style token it learned in training. It is repetitive, and we say so; its melodies sit in a proper vocal register for the first time, and generation is from scratch, with no data seeding. The style dropdown now only advertises what a loaded model actually understands.

30 Aug 2026

The bar 31 problem

The first listener of an op. 9 melody noticed it began with thirty silent bars. The model had faithfully learned instrumental intros: melodies extracted from real songs keep their original timestamps, so a vocal entering after an intro carried its dead air into training. Nearly three quarters of the corpus had the artifact. Every melody is now shifted left by whole bars, preserving pickup phrasing, and future studies learn from music that starts when the file does.

29 Aug 2026

A corpus of true melodies

The ninth study's training data went from 275 full arrangements to 18,758 single-voice melodies: POP909's human-labeled melody tracks, 5,981 Wikifonia lead sheets, and 11,870 lines extracted from the Lakh MIDI dataset by a skyline algorithm validated against human annotations (mean F1 0.869, roughly one accompaniment leak per nine hundred songs). Each melody carries a genre tag resolved from MusicBrainz and Last.fm metadata.

29 Aug 2026

The Melograph gets its name

The site formerly known as AI Melody Generator is now The Melograph, after the nineteenth century instrument built to capture melodies as they were played. New domain, new engraved look, and a working monograph describing the method behind all nine studies.