Question 1

What does jev-riffs do end to end?

Accepted Answer

It parses a MIDI file into note events, picks a melody track, converts it to an interval token string, mines repeating units in code, and has Jev grade every candidate's musical significance in one batched request. A web UI then draws the piano roll and plays each result.

Question 2

Why tokenize intervals instead of notes?

Accepted Answer

Two properties matter: the alphabet stays small (mostly plus or minus 12 semitones), and the encoding is transposition-invariant, so a riff restated a fourth higher produces the identical token substring and the miner sees through key changes for free.

Question 3

How are candidate motifs found?

Accepted Answer

By counting, not by model judgment: the miner enumerates windows, counts non-overlapping full occurrences with a greedy scan, keeps candidates seen at least twice, prunes shadows, periodic units, rotation echoes, and superstrings, then ranks by coverage and caps the list at the top 20.

Question 4

What does Jev add to the pipeline?

Accepted Answer

Significance grading on four ordered levels: 0 incidental overlap, 1 recurring figure, 2 recurring phrase, 3 signature theme. The README explains the binary is-this-a-motif question saturated on music, with Jev saying yes to all 20 candidates in the first live run, so the graded version replaced it.

Question 5

How did the Swan Lake test go?

Accepted Answer

The README reports the ripper pulled the Swan Lake theme out of a 17-track orchestral MIDI as its top-ranked motif, in one batched API call, in roughly 600 ms. That is the project's one documented proof case, not a benchmark suite.

Question 6

What can I do with the output?

Accepted Answer

The web UI draws the piano roll, highlights where each motif lives, plays the full passage on click, and exports WAV or JSON. The README links a demo video that loads Swan Lake, rips it, and plays the theme when you click a motif.

Question 7

What are the limits of the ripper?

Accepted Answer

It uses a melody heuristic - the densest track after dropping drums, with chords collapsed to the top note - and deliberately ignores durations and controllers, so it needs pitch order rather than performance data. It needs a Jev API key, and the grading budget and chunking parameters bound large pieces.

Jev Riffs

Highlights

Questions about this entry

From the Jev guides

From the Jev AI model community