Clawgrip sent me a UM3482A, a melody generator chip used in a Tetris bootleg.  The datasheet says it has a 512 note memory, and each note is 7 bits long.  The chip can address a maximum of 16 songs, but the 3482 only has 12 songs.

The datasheet says there are "8 beats selectable" and "14 tones selectable", which I assume means 8 note durations (using 3 of the 7 bits) and 14 tones (which would use the remaining 4 bits). The 2 "missing tones" could be pause and end of song flags.

I decapped the chip and removed the top metal layer.  The ROM array is easy to spot in the upper middle of the chip.  It has 64 rows of 56 columns, for a total of 3584 bits, which matches the datasheet's 512 7-bit notes.
The 56 columns are made up of 7 groups of 8, so I assume each group of 8 columns is one bit position, and each note gets 1 bit from each group of columns, from the same row and same sub-column in each group.
I think the notes start at the top row, and the following notes are in lower rows, until the 64th note in the bottom row.  Then the next note starts over at the top row, but in a different sub-column.

I'm not sure what order the groups are in; group 1 could be the msb or lsb, or they might not even be sequential.  I don't think the sub-columns are sequential; sub-column 4 in groups 1, 6 and 7 is all 1s, and in groups 2, 3, 4 and 5 is all 0s.
And the last 23 rows of sub-column 3 have the same pattern.  I think that's filler after the last song.

There are 3 other smaller ROM arrays on the die below the main ROM.  The leftmost one has 8 7-bit values, the middle one has 16 9-bit values, and the rightmost one has 16 7-bit values.  Since there are a max of 16 songs, and 512 notes
require 9 address bits, I think the middle ROM has the starting addresses for each song.  The bottom value is all zeros, and each one is larger than the one below (msb on the right).  Even though there are only 12 songs, the top 4 words are populated-
maybe this is garbage, or perhaps there are 4 more songs in ROM that can't normally be played.  The rightmost ROM also has 16 values, so I assume that's one per song.  Maybe the tempo for each song?

I connected the chip inputs to play each song once then stop, provided a 100KHz clock input, and captured the raw digital output on the MTO pin with no envelope circuit connected.  I also converted that output to a WAV file.
But it seems like all the songs together use more than 512 notes- maybe as many as 800- so I don't know what that means.

I tried to match notes in the ROM array to what are played, but didn't have any luck.

I purchased a UM3481A to compare.  It has 8 Christmas songs.  The die looks identical except for the note ROM and 2 of the other ROMs.  The leftmost ROM is identical to the 3482.

um3481araw.bin and um3482araw.bin are the visual transcriptions of the ROM arrays.

um3481a_all_1.logicdata and um3482a_all_1.logicdata are the Saleae Logic captures of all songs playing on each chip.  You can download their software free to load the files.

um3481a_all_1.bin and um3482a_all_1.bin are those files converted to binary dumps.  um3481a_all_1.wav and um3482a_all_1.wav are the files converted to WAVs.