Need help for research project (Call for participants)

[realfolkcode]

Hi all!

I am a music enthusiast and a PhD student doing research in the intersection of music and machine learning. My research project focuses on the evaluation of generative piano performance models. In particular, the scope of such models is within the realms of classical music repertoire.

The motivation behind these models is to make music performance accessible (e.g., when someone faces difficulties learning music theory or struggles due to physical limitations). Such models show some promise in following emotional and interpretive cues from humans, which can be input via text, voice, or gestures.

I am conducting a listening test to evaluate the performance of MIDI piano models. Participants are asked to rate the naturalness and expression of short excerpts of classical piano music. The goal of this experiment is to investigate how well our proposed evaluation metrics align with human perception.

Link to the test: https://test.piano-eval.com/ad?recruiter=generic&generate_tokens=1&mode=live
Format: Online (via web browser) Note that Safari is not supported
Duration: Approximately 15-20 minutes
Preferred conditions: A quiet environment and headphones

Apologies if this topic is outside the scope of this forum. I would greatly appreciate your participation in this test! If you have any questions, please feel free to reach out and chat!

Thanks!

[essence]

Interesting. I did the test. I identified them all as AI generated or else a very poor pianist. None was the slightest bit musical.

I recognised most of the extracts, but it may be a good idea to say what the pieces are at the end.

[realfolkcode]

Interesting. I did the test. I identified them all as AI generated or else a very poor pianist. None was the slightest bit musical.

I recognised most of the extracts, but it may be a good idea to say what the pieces are at the end.

Thank you for your help and time! Yes, you are correct, all the performances are AI generated, and there are no human performances in the test. I agree it's a missed opportunity to display the pieces after the test completion.

[essence]

Might have been an idea to include human performances to see if anybody can identify them?

[keypeg]

I am a music enthusiast and a PhD student doing research in the intersection of music and machine learning.

A first question coming to mind is whether as a music enthusiast you also play an instrument, and in particular, the piano.  That may be relevant.

The motivation behind these models is to make music performance accessible (e.g., when someone faces difficulties learning music theory or struggles due to physical limitations). Such models show some promise in following emotional and interpretive cues from humans, which can be input via text, voice, or gestures.

That part doesn't make sense to me, especially after having listened to the examples. It is also why I asked whether you have experience playing piano yourself.  If a student did need examples, there are many real and excellent examples on the Internet, and the old fashioned world of purchasing recordings.

Your models did not follow any kind of emotional or interpretive cue.  All of them had metronomic timing like if I write music into an app and it plays it back literally.  Many of the differences sounded like somebody didn't know how to use the pedal, or had a malfunctioning pedal.  I had to stop listening because it was unpleasant to listen to.

..... our proposed evaluation metrics align with human perception.

I'm curious what kinds of metrics those might be, and whether those involved in the project are musicians.  If it is to help learners, is anyone trained in teaching?

[realfolkcode]

A first question coming to mind is whether as a music enthusiast you also play an instrument, and in particular, the piano.  That may be relevant.

That part doesn't make sense to me, especially after having listened to the examples. It is also why I asked whether you have experience playing piano yourself.  If a student did need examples, there are many real and excellent examples on the Internet, and the old fashioned world of purchasing recordings.

Your models did not follow any kind of emotional or interpretive cue.  All of them had metronomic timing like if I write music into an app and it plays it back literally.  Many of the differences sounded like somebody didn't know how to use the pedal, or had a malfunctioning pedal.  I had to stop listening because it was unpleasant to listen to.

I'm curious what kinds of metrics those might be, and whether those involved in the project are musicians.  If it is to help learners, is anyone trained in teaching?

Hi! First of all, thank you for the extended feedback!

Regarding the first question, I study and play music as a hobby. My background is over ten years of playing a guitar (no particular genre or style, ranging from classical to pop and jazz). As for piano, my skills are much more poor, and I used to sketch some self composed ideas using a digital keyboard in the past.

I cannot really disclose my personal opinion the quality of these examples, as the listening test is still going. But it is your right to think the provided examples are poor, and I am grateful you partially completed the test and provided an extended critique.

I apologize for the possible misunderstanding. The provided audio examples were generated without any emotional or stylistic prompt. The main goal of this experiment is to evaluate the base performance of the models, i.e., without such prompts. None of the models were designed as part of this research project but we used the models from existing papers in the field of generative performance rendering.

The metrics are rooted in mathematical theory of machine learning. Unfortunately, for now, I can not disclose the details, as the upcoming paper will be subject to peer review. None of my coauthors perform music professionally. However, my research supervisor is a great admirer of classical music and has a deep knowledge in this field.

[radjabov]

I have already finish your test it was longer than i expected. So similar performances hard to evaluate all that questions, in general very poor human like recordings maybe the poor quality of digitize piano's sounds didn't help so much.

[essence]

I have to say that your research project desperately needs any advanced pianist or piano teacher to help you.  Being a guitar player is far from sufficient.

What is the point of demonstrating AI sucks, and you don;t have enough experience to see how it might be improved?