Sight Reading app recommendations

Victoria Chan Victoria Chan 

Noel Nguyen I already have. It was helpful, but my issue isn't lack of knowledge as much as complete lack of talent (and the fact that I have dysmusia/musical dyslexia). 

Maria Interesting diagnoses that I have never heard of! At the very least you are medically literate.

Noel Nguyen I think dysmusia is more common than one may think, especially in those with absolute pitch who can work around it by playing by ear. 

Erik Satie almost certainly had dyslexia and dysmusia, and Chopin and Beethoven probably had dyslexia (but not dysmusia). 

Maria From a quick survey of the literature I don't think most doctors would know about dysmusia. But then again the medical field is known for not being so quick to recognize new entities. Even focal dystonia is considered fairly new.😬

Noel Nguyen No, it is not well-known outside of "dyslexic students struggle more than average with sight-reading." 

I am not formally diagnosed but my teachers and I am 99.9% sure I have it, and my experience of reading sheet music is nearly identical to dyslexic people's descriptions of their experiences of reading words. 

Maria One thing to keep in mind is that the medical field being so slow to recognize new clinical entities is not necessarily a bad thing. It prevents potential harm that could come from hastily accepting proposed treatments that may turn out to be harmful in the long run... That's one of the reasons it always takes a very long time for notions to go from alternative medicine to mainstream medicine, if ever.

But enough academics for tonight! I'm going back to the piano now to refill my mind with good stuff!

Bret Cameron Very cool! I will try it.

I have been talking to Claude AI about building a multi-pass post hoc analysis tool which will give me feedback on rhythm, hand alignment, and missing or wrong notes. 

I think a post hoc tool makes the analysis problem easier. Any analysis of interpretation is outside of the scope of the project.

I would like your option about building such a tool. 

I have an acoustic piano, a Yamaha DYUS5 that can output MIDI as I play acoustically. I can also record can play back mechanically.
 

Larry K That's great, hope you enjoy it - and let me know if you have any thoughts!

I think using Claude to build a "post-match review" is a great idea, though you may find SightReader already does similar to what you're looking for. 

After playing, you can get feedback on pitch, timing and dynamics (see screenshots below). It's also possible to upload your open pieces, if you have (or can get) a MusicXML version of them.

Bret Cameron I can get MusicXML for any piece I am working on by writing it up using the StaffPad app with an Apple pencil on my iPad Pro and exporting in MusicXML format.

Claude says I get this from post hoc analysis,

”

What "post hoc + piano + intermediate/advanced" changes

• Global alignment is possible, so two or three passes beat any online approach. Early passes build a tempo model, and later passes use it to resolve ambiguities.
• Rubato is expected. At this level, "tempo error" mostly can't mean "differs from the marked BPM." I'd define two modes:
  • Self-consistency (default): flag unintended unsteadiness, such as rushing, stalls, or uneven subdivisions.
  • Target tempo: compare against the marked or user-supplied BPM (useful for metronome practice).
• Precision beats recall. Advanced players quickly ignore a tool that cries wolf, so every flag carries a confidence score, and ambiguous regions go into an "uncertain" bucket instead of being flagged.
• Piano-specific hazards are chord spread, hands that aren't together, ornaments, and repeated figuration (an Alberti bass makes a missed note alignable almost anywhere).

Pipeline: three passes

1. Coarse alignment. Pitch-dominant matching with a loose time prior gives a rough path and a first tempo curve.
2. Refined alignment. Re-run the DP with the tempo curve as the time prior, which fixes ambiguity in repeated patterns. Iterate once or twice.
3. Note-level matching and classification. Resolve the matched slices into note-level results, then run the timing analysis on the matched notes.”

Are you doing post hoc analysis or live analysis? 

Larry K This is very interesting.

For SightReader it's a mixture of live analysis (for pitch and rhythm) and post hoc analysis.

In the post hoc section I have:

• Dynamics (this could be live, but isn't because it's too much visual clutter alongside pitch and rhythm feedback)
• Repeated patterns in the playing (such as a repeatedly misplayed pitch). 

There is maybe room to do more. Right now, I only show what can be measured with a high degree of confidence. So for example dynamic feedback only identifies clear dynamic mistakes, rather than anything that could be more subtle.

Any rubato would also currently be marked as a mistake, but I think for a tool like this which is training reading ability, it's better to mark against strict time than being overly generous!

Bret Cameron I agree that strict time is better than assuming rubato. 
 

Here is my conversation with Claude, 

https://claude.ai/share/0c68a8eb-3983-42df-b77f-8c8b9744b1ec

I have a software background but not a music analysis background.

I pushed Claude to write some code in Smalltalk, an environment that allows for live programming. Any method on any object can be halted in the debugger, rewritten, and execution continued from that stack frame.

This had the unfortunate effect of making it more difficult to use some Python libraries. If you want, I can ask Claude to generate Python instead. 

What did you write your app in?

Let me know if any parts of that conversation are useful to you. There are some good sources for music analysis in the thread, including a paper by Nakamura.

Maybe you could introduce an offline matching mode.