Hi, Kevin! Yes, I used Perplexity to attempt to act as an OMR and mentioned my experience in a prior message in this thread, but I’ll recount some bits here.
Even with a reasonably straightforward score and an extra simplified goal of basically only recognizing staves, time signature, measures, note durations and pitches, it fell quite short. It did recognize the two staves of the piano and the general trend of the notes, but beyond that, nothing was correct. The code it generated, however, was complete and structured well and compiled without errors. I tried the same exercise with DeepSeek and it failed even more spectacularly. It also generated working, well-structured code, but the content was 100% wrong except for recognizing it as a two-staff piano score. It does feel like it’s only a matter of time before at least one of the tools is able to get this job done, but there is still some ways to go. Here is the prompt I ended with, in case anyone wants to play around with it: %< - - - You are an expert in LilyPond notation and optical music recognition. Given the attached sheet music image, transcribe ONLY the structural and foundational musical elements into clean, efficient LilyPond code. **PRIORITIZE ACCURATELY CAPTURING:** 1. Score structure (number of staves, instruments, staff grouping with braces/brackets) 2. Clefs (type and position changes) 3. Key signatures (exact sharps/flats) 4. Time signatures (including changes) 5. Note pitches and durations 6. Rests (type and duration) 7. Ties (between notes) 8. Basic voicing (voice separation where visible) **EXPLICITLY IGNORE AND OMIT:** - Articulations (staccato, accents, tenuto, etc.) - Dynamics (p, f, crescendo hairpins, etc.) - Slurs, phrase marks, and ties (except note-to-note ties) - Text annotations, lyrics, rehearsal marks - Ornamentation (trills, turns, mordents) - Fingerings - Special techniques (pizz., arco, con sordino, etc.) - Exact visual positioning/coordinates - Page layout details **WORKFLOW:** 1. First, analyze the image and provide a brief structural summary: "3-staff piano score, treble/bass clef, 4/4 time, C major, 24 measures..." 2. Generate clean LilyPond code with: - Minimal necessary header - Correct staff/voice structure - Notes, rests, and ties ONLY - No layout overrides for visual perfection - Clear comments for each section 3. Use standard LilyPond conventions (automatic stem directions, standard beaming) 4. If uncertain about a note or rhythm, mark it with `%%% QUESTION: [description]` in the code 5. Keep the code readable and well-organized for easy manual enhancement Generate code that represents the musical skeleton accurately, not a visual replica. %< - - - HTH, Abraham On Tue, Jan 13, 2026 at 5:17 AM Kevin Cole <[email protected]> wrote: > Has anyone tried Perplexity.ai? > > In my limited experience with both LilyPond and AI's in general, I > have not had the need to try combining the two. However, of the free > AIs I' fooled with, I'm liking Perplexity the best -- though mainly as > a search engine that actually searches for what I ask it to, without > giving me a lot of completely irrelevant results that ignore most of > my search terms. When I DO actually ask it for something more > "how-to", I love that it cites its sources: One can then (a) verify if > the sources are legitimate sources of good information and (b) > determine how badly Perplexity mangled the information it was mashing > up. > >
