ðé³å£°ããïŒæ¥ïŒè±ïŒïŒãæªæ¥ã®é³æ¥œãAIããããã¹ã®æ¥œèãæžãæä»£ãžïŒææ°ç ç©¶ãMulTTiPopãã培åºè§£èª¬
ð¥ æ¬æ¥ã®è«æãšããã«ã€ããŠã®åŠæ³ïŒæ¥æ¬èªçïŒ
ð
ð ã¿ã€ãã«ïŒãæªæ¥ã®é³æ¥œãAIããããã¹ã®æ¥œèãæžãæä»£ãžïŒææ°ç ç©¶ãMulTTiPopãã培åºè§£èª¬
ð æ¬æïŒæ¥æ¬èªïŒ
ã¿ããªãããã«ã¡ã¯ãŒïŒ
ããããäžã®å
ãã£ãä»®ã ãã
ããã仿¥ãå
æ°ã«ãã£ãŠã¿ããïŒ
仿¥ã¯ã2026幎7æ13æ¥æææ¥ã ãã
ãã®æéã¯ãããããã€ã³ã¿ãŒãããã®ã¢ãŒã«ã€ãã§èŠã€ããã
ãããããããããªãã¬ã³ãè«æã玹ä»ããŠãããã
仿¥ã®ããŒãã¯ãã¿ããªã倧奜ããªæ©æ¢°åŠç¿ã ãã
AIã鳿¥œãèŽããŠã
ãŸãã§éæ³ã®ããã«æ¥œèãæžãèµ·ãããŠãããã
ãããªå€¢ã®ãããªã話ãªãã ã
ã¡ãã£ãšé£ãããã«èããããããããªããã©ã倧äžå€«ã
ãããããããããããŒã説æãããããæåŸãŸã§èããŠãã£ãŠãã
ãŸãã¯ã仿¥ç޹ä»ããè«æã®ã¿ã€ãã«ãšURLããç¥ãããããã
ã¿ã€ãã«ã¯ã
MulTTiPop: A MULTITRACK TRANSCRIPTION DATASET FOR POP MUSIC
URLã¯
https://arxiv.org/abs/2607.08756v1
ã ããã¿ã€ãã«ãã¡ãã£ãšãã£ãããããïŒ
ããŠããŠããã®è«æãã©ããªåé¡ã解決ããããšããŠããã®ãã
詳ããã話ãããŠãããã
ã¿ããªã¯ã奜ããªãããã¹ã®æ²ãèŽããŠããŠã
ãã®æ²ã®ãã¢ãã®æ¥œèãæ¬²ãããªãšãã
ããŒã¹ã¯ã©ããã£ãŠåŒŸããŠããã®ããªãã£ãŠæã£ãããšã¯ãªãããªïŒ
ãããã³ã³ãã¥ãŒã¿ãŒã«èªåã§ãã£ãŠãããæè¡ãã
èªå鳿¥œæ¡èãè±èªã®é æåããšã£ãŠAMTãšåŒã¶ãã ã
æè¿ã®æ©æ¢°åŠç¿ãã€ãŸãAIã®é²æ©ã§ã
ãã®AMTã¯ãããè³¢ããªã£ãŠãããã ãã
ã§ãããå®ã¯å€§ããªåŒ±ç¹ããã£ããã ã
ããã¯ãäžäººã§åŒŸããŠãããã¢ãã®æ²ãã
ã¯ã©ã·ãã¯é³æ¥œã®æ¥œèãäœãã®ã¯åŸæãªãã ãã©ã
ç§ãã¡ãæ®æ®µããèŽããããªã
ããŒã«ã«ãã®ã¿ãŒããã©ã ãããããéãªã£ãŠããããããã¥ãŒãžãã¯ã ãšã
æ¥ã«èŠæã«ãªã£ãŠããŸãããšãªãã ã
AIã鳿¥œã®å
šäœåãæããããªããŠã
æ··ä¹±ãã¡ãããã ããã
ãã£ãšãã©ãããŠãããªããšãèµ·ããã®ããªïŒ
å®ã¯ãAIãè³¢ãè²ãŠãããã«ã¯ãæç§æžãšãªãããŒã¿ãå¿
èŠãªãã ã
ãããŸã§ã®æç§æžã¯ãã³ã³ãã¥ãŒã¿ãŒã§äœã£ãåœç©ã®é³æ¥œããŒã¿ã ã£ããã
ã¯ã©ã·ãã¯é³æ¥œã°ããã ã£ãããããã ããã
ã ãããæ¬ç©ã®ãããã¹ã®è€éãªé³ãèŽãåãããã¹ãããããšã
è¯ãç¹æ°ãåããªãã£ããã ã
ããã§ããã®è«æã®ç ç©¶è
ãã¡ã¯èãããã ã
æ¬ç©ã®ããããã¥ãŒãžãã¯ãšã
ãã®æ£ããæ¥œèãã»ããã«ãªã£ãã
æ°ãããã¹ãçšã®æç§æžãäœããã£ãŠãã
ããããã¿ã€ãã«ã«ããã£ãMulTTiPopãšããããŒã¿ã»ãããªãã ã
ãã®MulTTiPopã¯ããã ã®ããŒã¿ãããªããã ãã
ã€ã³ã¿ãŒãããã§ã¿ããªãèŠãŠããYouTubeã®é³æ¥œåç»ããã
æ¬ç©ã®ãããã¹ã®é³æºãéããŠããŠãããã ã
ãããŠããã®é³æºã«åãããŠã
MIDIãšåŒã°ããã³ã³ãã¥ãŒã¿ãŒçšã®æ¥œèããŒã¿ãã
ãŽã£ãããšã¿ã€ãã³ã°ãåãããŠãã£ã€ãããã ãã
ãããäœãã®ã¯ãæ¬åœã«å€§å€ãªäœæ¥ã ã£ãã¿ããã
æåã¯ããã°ã©ã ã䜿ã£ãŠèªåã§ã¿ã€ãã³ã°ãåãããããšãããã ãã©ã
ããã ãã ãšãããŠããŸãããšãå€ãã£ããã ã£ãŠã
ã ããæåŸã¯ã人éã®ã¹ã¿ãããããã¡ãäžã€äžã€è³ã§èŽããŠã
äžçªãŽã£ããåãã¿ã€ãã³ã°ãæäœæ¥ã§éžãã ãã ã
ãããã£ãŠãçŽ3.5æéåã572åã®ã
é«å質ãªãã¹ãããŒã¿ãäœãäžãããã ãã
ããããããã®è«æã§ã¯ãä»ã®æè¡ãšã®æ¯èŒãè¡ã£ãŠãããã ã
ãããŸã§ã«ããRWC-Popãšããã
ãããã¹çšã®ããŒã¿ã»ããããã£ããã ãã©ã
ããã¯ç ç©¶ã®ããã«ç¹å¥ã«äœãããæ²ã°ããã§ã
äžã®äžã§æ¬åœã«ãããããŠããæ²ãšã¯å°ãéã£ããã ã
ã§ãMulTTiPopã¯ãå®éã«ããããã倿§ãªãžã£ã³ã«ã®æ²ãå«ãã§ããããã
ããçŸå®ã®äžçã«è¿ããã ãã
ããã«ãçŸåšæãåªç§ãšèšãããŠããã
MT3ããYourMT3+ãšãããæå
端ã®AIã¢ãã«ã«ã
ãã®MulTTiPopã®ãã¹ããè§£ãããŠã¿ããã ã
ããããããäžçªæçžŸã®è¯ãã£ãã¢ãã«ã§ãã
æ£è§£çããçŽã38%ãããªãã£ããã ã£ãŠïŒ
ããã¯ãAIã«ãšã£ãŠã¯ãŸã ãŸã é£ããåé¡ã§ã
ãããããã£ãšãã£ãšæé·ã§ããäœå°ãããã
ãšããããšã蚌æããçŽ æŽãããçºèŠãªãã ãã
ããŠããã®æè¡ããã£ãšé²åããŠã
å®ç§ã«ãããã¹ã®æ¥œèãäœããããã«ãªã£ããã
ç§ãã¡ã®æ¥åžžç掻ã«ã©ããªåœ±é¿ãäžãããšæãïŒ
å
·äœçã«äžã€ã®å¿çšäŸã玹ä»ãããã
äžã€ç®ã¯ã楜åšã®ç·Žç¿ããã£ãšæ¥œãããªãããšã ãã
ã¿ããªãYouTubeã鳿¥œã¢ããªã§èŽããŠããã
ãæ°ã«å
¥ãã®ææ°ãããæ²ãããããã
ãã®æ²ã楜åšã§åŒŸããããšæã£ãæã
AIãããã«ãã®æ²ãèŽãåã£ãŠã
ããªãå°çšã®æ¥œèãç»é¢ã«åºããŠãããããã«ãªããã ã
ãã¢ãã§ãã®ã¿ãŒã§ããèªåã®å¥œããªæ¥œåšã®ããŒãã ããæãåºããŠã
ãã£ããåçããªããç·Žç¿ã§ããããã«ãªããã ãã
äºã€ç®ã¯ã誰ã§ãç°¡åã«é³æ¥œãããã¥ãŒãµãŒã«ãªããããšã ãã
倧奜ããªæ²ã®ããã©ã ã®é³ã ããå¥ã®ãªãºã ã«å€ããŠã¿ãããªãšãã
ããŒã«ã«ã®å£°ã ããåãåºããŠãå¥ã®æ²ãšæ··ããŠã¿ãããªã£ãŠæã£ãæã«ã
ãã®æè¡ã䜿ãã°ãè€éã«æ··ããåã£ãé³ã®äžããã
ç¹å®ã®æ¥œåšã®é³ã ãããããã«åãåºãããšãã§ãããã ã
ããã䜿ãã°ããªãªãžãã«ã®ãªããã¯ã¹ãäœã£ããã
æ°ãã鳿¥œãçã¿åºãã®ããéæ³ã®ããã«ç°¡åã«ãªããã ãã
äžã€ç®ã¯ããã£ãšè³¢ã鳿¥œæ€çŽ¢ã·ã¹ãã ãã§ããããšã ãã
ä»ã¯æ²ã®ã¿ã€ãã«ãææã®ååã§æ€çŽ¢ããããã
ã§ãããã®æè¡ãé²åããã°ã
ãã®æ²ã®ããã®ã®ã¿ãŒã®ãã¬ãŒãºããã£ãããæ²ããšãã
ããŒã¹ããããªæãã§åãæ²ãæ¢ããŠã
ã¿ããã«ã鳿¥œã®äžèº«ãã®ãã®ã§æ€çŽ¢ã§ããããã«ãªããã ã
ããããã°ãä»ãŸã§åºäŒããªãã£ããããªã
ããªãã®å¥œã¿ã«ãŽã£ããã®æ²ããAIãèŠã€ããŠãããããã«ãªããã ãã
ã©ãããªïŒ
AIã鳿¥œãçè§£ããŠã楜èãæžããŠãããæªæ¥ã
ãªãã ããšãŠãããããããŠããªãããªïŒ
ãã®MulTTiPopãšããç ç©¶ã¯ã
ãããªæ¥œããæªæ¥ãå®çŸããããã®ã
ãšã£ãŠã倧åãªç¬¬äžæ©ãªãã ãã
鳿¥œãšãã¯ãããžãŒã亀ãããšããã«ã¯ã
ãã€ãæ°ããçºèŠããã£ã±ããããã
ããããããã£ãšãã£ãšæ°ããæè¡ãå匷ããŠã
ã¿ããªã«é¢çœãã話ãå±ããŠãããããªã
仿¥ã¯ãæ©æ¢°åŠç¿ã䜿ã£ãŠããããã¥ãŒãžãã¯ã®æ¥œèãäœãã
ãããªçŽ æµãªç ç©¶ã«ã€ããŠç޹ä»ããŸããã
æåŸãŸã§èŽããŠãããŠãæ¬åœã«ããããšãïŒ
ãŸã次åã®æŸéã§äŒãããã
ã°ãã°ãŒãïŒ
ð The Paper and Some Imagination (English)
ð
ð TitleïŒ How AI Learns Pop Music: The MulTTiPop Breakthrough
ð Summary (English)
Hello everyone, and welcome back to our fun little classroom on the radio.
I am your host, ichino-ani,
and I am so excited to dive into another fascinating lesson with all of you today.
Today is Monday, July thirteenth, twenty twenty six,
and I have a really amazing topic to share with you all.
Are you ready to put on your thinking caps?
Well, today we are looking into the wonderful world of machine learning,
and specifically, how artificial intelligence listens to music.
The title of the paper we are looking at today is called,
MulTTiPop A MULTITRACK TRANSCRIPTION DATASET FOR POP MUSIC.
And the URL for this paper is,
https://arxiv.org/abs/2607.08756v1
Ah, yes, that is quite a mouthful,
but I promise the ideas inside are super exciting.
So, what is the big problem this paper is trying to solve?
Um, imagine you are listening to your absolute favorite pop song on the radio.
You can probably hear the singer singing,
the drums beating a cool rhythm,
maybe a sparkly synthesizer,
and a deep bass guitar thumping away in the background.
As humans, we can naturally tell all these sounds apart,
but for a computer, this is like listening to a giant, messy soup of noise.
Right now, scientists have built AI models that are pretty good at writing down sheet music if they are only listening to one instrument,
like a solo piano playing a beautiful classical piece.
But when it comes to real world pop music,
with all its loud, layered instruments and commercial production tricks,
the AI gets completely confused and jumbles everything up.
The problem is that the computer models simply do not have a good textbook to learn from.
The older textbooks, or datasets, as scientists call them,
usually only feature classical music,
or they use fake, robotic synthesizer sounds that do not sound like real human musicians at all.
Some datasets do have real pop music,
but they only teach the computer how to find the main melody,
ignoring all the other wonderful instruments playing in the background.
Well, the brilliant researchers at Carnegie Mellon University decided to fix this big problem.
They created MulTTiPop,
which is a brand new dataset designed to be the ultimate textbook for teaching AI how to transcribe multitrack pop music.
They gathered five hundred and seventy two segments of real, commercially produced pop music from YouTube,
spanning all the way from the nineteen thirties to the two thousands.
Then, they matched these real audio clips with highly detailed MIDI files.
A MIDI file is basically digital sheet music that tells a computer exactly what note to play,
when to play it,
and on what instrument.
To make this work, they had to be very clever.
They used computer algorithms to match the song titles and artists,
and then they used a special beat tracking method to gently stretch and squeeze the digital sheet music,
so that it perfectly matched the tiny changes in tempo of the real human performance.
Finally, they even had real human students check the work to make sure the digital notes lined up perfectly with the YouTube audio.
When they tested some of the best AI models in the world on this new MulTTiPop dataset,
the models did not do very well at all.
The very best model only achieved thirty eight percent accuracy in finding the right notes and instruments.
This shows us that there is a huge amount of room for improvement,
and MulTTiPop is going to be the perfect challenge to help scientists build smarter AI in the future.
Now, you might be wondering,
how does this technology actually help us in our everyday lives?
Let me give you three really cool examples of how this could change the world.
First, think about music education and learning to play an instrument.
If an AI can perfectly listen to any pop song and write down the exact notes for every single instrument,
it would make learning music so much easier for kids and adults everywhere.
Imagine you hear a brand new song on the radio,
and you really want to learn how to play that exact drum beat or that specific guitar solo.
Instead of searching the internet for sheet music that might not even exist,
the AI could instantly generate the perfect sheet music just for you.
You could see exactly what the bass player is doing,
and what the keyboard player is doing,
helping you learn and practice your favorite songs almost instantly.
Second, this technology would be absolutely amazing for music production and remixing.
Music producers love to take old songs and give them a fresh new twist,
but it is really hard to separate all the instruments once they are mixed together into one audio file.
With a super smart AI trained on MulTTiPop,
a producer could feed an old pop song into the computer,
and the AI would give them the digital MIDI notes for every single part.
The producer could then take the exact melody the singer sang,
and easily have a computer synthesizer play those exact same notes instead.
Or they could take a classic bassline from the nineteen seventies,
and completely change the sound of the instrument to make a futuristic dance track.
It gives artists a magical tool to take apart a song like puzzle pieces,
and put it back together in totally new and creative ways.
Third, think about preserving music history and helping with copyright protection.
Every day, thousands of new songs are uploaded to the internet,
and sometimes artists accidentally write a melody that sounds exactly like a famous song from thirty years ago.
If we have AI that can perfectly transcribe the multitrack details of every song ever recorded,
we could build a massive, searchable library of musical DNA.
If someone claims that a new hit song stole their guitar riff,
the computer could instantly compare the digital transcriptions and see exactly how similar they really are.
This helps protect artists and ensures that everyone gets the proper credit for their hard work and creativity.
When we compare this new MulTTiPop dataset to older technologies,
the difference is truly like night and day.
For example, an older dataset called MAESTRO is very famous,
but it only contains solo classical piano music.
That is like trying to learn how to paint a colorful jungle by only using a grey pencil.
Another dataset called Slakh2100 has lots of different instruments,
but all the audio is synthesized by a computer,
which means it sounds flat and robotic.
Real pop music has singers with unique voices,
guitars playing through messy distortion pedals,
and drummers hitting cymbals with different amounts of force.
By bringing real, messy, beautiful commercial pop music into the training room,
MulTTiPop forces the AI to learn how to hear the world the same way we do.
It is a big, difficult challenge,
but it is exactly what we need to make computers truly understand the magic of music.
Ah, we have learned so much today.
It is so wonderful to see how computer scientists are working hard to bridge the gap between cold algorithms and the deeply emotional world of popular music.
I hope you all enjoyed today's lesson,
and maybe the next time you listen to your favorite song,
you will try to listen really closely to every single instrument,
just like our AI friends are learning to do.
Thank you so much for joining me in our radio classroom today.
Have a fantastic day,
and keep exploring the wonderful world of science and music.
Goodbye everyone!
ðïž ã³ã¡ã³ã
æåŸãŸã§èªãã§ãããŠæ¬åœã«ããããšãïŒïŒ
ãã€ãã©ãããããŸã話ããªããïŒããããããããããïŒ
åçãªã¹ãã§ãŸãšããŠãããããæ°ãåãããèŽããŠã¿ãŠãïŒ
æ¥æ¬èªã¯ð
è±èªã¯ð
äœèšã£ãŠããåãããªããã©ãèŽããŠããåããããã«ãªãããïŒïŒ
åãããªããŠãåå®åã®ä»£ããã«èŽããŠã¿ãŠãïŒ
Original paper link: ð
ãé¢é£ããŒã¯ãŒãã
#AI #æ©æ¢°åŠç¿ #鳿¥œ #楜è #èªåæ¡è #MulTTiPop #AMT #YouTube #MIDI #ãã¯ãããžãŒ #æªæ¥ #鳿¥œå¶äœ #DTM #ææ°ç ç©¶ #AIMusic #MachineLearning #PopMusic #MusicTranscription #MulTTiPop #ArtificialIntelligence #CarnegieMellon #MusicTech #DataScience #AudioAnalysis #HowAILearns
