ðé³å£°ããïŒæ¥ïŒè±ïŒïŒAIããã³ãã¡ã³ããŒã«ïŒæ¬¡äžä»£ãªã¢ã«ã¿ã€ã 鳿¥œçæãLive Music Modelsã解説
ð¥ æ¬æ¥ã®è«æãšããã«ã€ããŠã®åŠæ³ïŒæ¥æ¬èªçïŒ
ð
ð ã¿ã€ãã«ïŒAIããã³ãã¡ã³ããŒã«ïŒæ¬¡äžä»£ãªã¢ã«ã¿ã€ã 鳿¥œçæãLive Music Modelsã解説
ð æ¬æïŒæ¥æ¬èªïŒ
ã©ãããŒãäºã®å ãã£ãä»®ã§ãã
2025幎8æ8æ¥éææ¥ã
é²ã²ãšã€ãªãé空ãåºãã£ãŠãããšãããªãŒããªããŠæããªããã
仿¥ãå§ããŠãããããªã
ããŠããŠã
ãã®ã©ãžãªã¯ããªã¬ãæ°ã«ãªã£ãã¢ãŒã«ã€ãã®è«æãã
ãããŒãã
ç¬ãèšã¿ããã«ç޹ä»ããŠããã
ã¡ãã£ãšå€ãã£ãçªçµã ãã
仿¥ã®ã«ããŽãªãŒã¯ã
ãµãŠã³ããé³é¿ã ãã
é³é¿ä¿¡å·ã®åŠçãšãã
åæã
åæã«é¢ããåéã
ãªã¬ãçµæ§å¥œããªãã ãããŒã
ãšããããã§ã
仿¥ç޹ä»ããã®ã¯ããã ãã
ã¿ã€ãã«ã¯ã
Live Music Models.
URLã¯
https://arxiv.org/abs/2508.04651v1
ã ãã
ã©ã€ã ãã¥ãŒãžã㯠ã¢ãã«ãºã
ã ã£ãŠã
ãªãã ãã
èãã ãã§ã¯ã¯ã¯ã¯ããååã ãã
ãã®è«æã解決ããããšããŠãåé¡ã£ãŠããã®ãã
ãŸãé¢çœããã ã
ä»ãŸã§ã®é³æ¥œãäœãAIã£ãŠã
ãªããŠãããã
äœæ²ã®ãé¡ããããæãã ã£ããã ããã
äŸãã°ã
80幎代颚ã®ãããã¹ã§ã
ã·ã³ã»ãµã€ã¶ãŒããã©ãã©ããæ²ãäœã£ãŠã
ã£ãŠãé¡ããããšã
AIãã
ããŒãã£ãŠèããŠã
ãã°ããåŸ
ã£ãåŸã«ã
ã¯ãã©ããã£ãŠã
宿ããæ²ãæž¡ããŠãããã
ããã¯ããã§ã
ãã¡ããããããã ãã©ã
ãã®è«æã§ã¯ã
鳿¥œã«ã¯äºã€ã®åŽé¢ãããã£ãŠèšã£ãŠããã ã
äžã€ã¯ã
ä»èšã£ãã¿ãããªã
é²é³ãããäœåãšããŠã®é³æ¥œã
ããã¯åè©ãšããŠã®é³æ¥œã ãã
ã§ãããäžã€ã
ã©ã€ãæŒå¥ã¿ããã«ã
ãã®å Žã§ãªã¢ã«ã¿ã€ã ã«äœéšãã鳿¥œãããããã
ããã¯ã
åè©ãšããŠã®é³æ¥œã
ãã¥ãŒãžããã³ã°ã£ãŠåŒã°ããŠããã ã£ãŠã
ããããªã¬ãä»ãã£ãŠãã®ã¯ã
ã©ãžãã³ã°ããªã
ããã
ã©ãžããã
ãã
ã©ã£ã¡ã§ããã£ãã
ããããããã
ãã®è«æãç®æããŠãã®ã¯ã
ãŸãã«ãã®åŸè
ã®ã
åè©ãšããŠã®é³æ¥œãªãã ã
ãŠãŒã¶ãŒãAIãšå¯Ÿè©±ããªããã
ãªã¢ã«ã¿ã€ã ã§ã
éåããããšãªã鳿¥œãäœãç¶ããã
ãããªæ°ããAIã
ã©ã€ã ãã¥ãŒãžã㯠ã¢ãã«ãææ¡ããŠããã ãã
ãããããã
ãŸãã§ã
AIããã³ãã¡ã³ããŒã®äžäººã«ãªãã¿ãããªæãããªã
ãããã
ã©ããã£ãŠãããªããšãå®çŸããŠãã®ãã
ãã®ä»çµã¿ãã
ã¡ãã£ãšã ãæ·±ãèŠãŠã¿ãããã
ãã®è«æã§ã¯ã
ããŒã³ã¿ ãªã¢ã«ã¿ã€ã ã£ãŠããã
誰ã§ã䜿ãããªãŒãã³ãªã¢ãã«ãšã
ã©ã€ã© ãªã¢ã«ã¿ã€ã ã£ãŠããã
ãã£ãšé«æ©èœãªAPIã¢ãã«ã®äºã€ã玹ä»ããŠããã ã
ä»åã¯ã
ç¹ã«ããŒã³ã¿ ãªã¢ã«ã¿ã€ã ã®æ¹ã«æ³šç®ããŠã¿ããããªã
ãã®ã¢ãã«ã®å¿èéšã¯ã
ã³ãŒãã㯠ã©ã³ã²ãŒãž ã¢ããªã³ã°ã£ãŠããæè¡ãªãã ã
ãŸãã
ã¹ãã¯ããã¹ããªãŒã ã£ãŠããã
ãã¥ãŒã©ã« ãªãŒãã£ãª ã³ãŒããã¯ã䜿ã£ãŠã
鳿¥œã®æ³¢åœ¢ããŒã¿ãã
èšèã¿ãããªã
ãã©ãã©ã®ããŒã¯ã³ã£ãŠãããã®ã«å€æãããã ã
é³ãããžã¿ã«ãªèšèã«ç¿»èš³ããæãã ãã
次ã«ã
ãã¥ãŒãžãã¯ã³ã«ã£ãŠããã¢ãã«ãã
ãŠãŒã¶ãŒããã®æç€ºãçè§£ãããã ã
äŸãã°ã
ããã¹ãã§ã
ãã¯ãã£ãœããã«ãŒãã®é³ã
ã£ãŠå
¥åãããã
åèã«ãªã鳿¥œãã¡ã€ã«ãæž¡ããããããšã
ãã®é³æ¥œã®ã¹ã¿ã€ã«ã衚ãç¹å¥ãªæ
å ±ã
ã¹ã¿ã€ã«ãšã³ããã£ã³ã°ã£ãŠããã®ãäœã£ãŠãããã
ãããŠæåŸã«ã
ãšã³ã³ãŒããŒãã³ãŒããŒæ§é ã®ã
ãã©ã³ã¹ãã©ãŒããŒã£ãŠããAIã¢ãã«ãç»å Žãããã ã
ãã®AIãã
éå»10ç§éã®é³æ¥œã®æµããšã
ãã£ãã®ã¹ã¿ã€ã«æ
å ±ãåãåã£ãŠã
ãããæ¬¡ã®2ç§ã¯ãããªæãããªã£ãŠã
æ°ãã鳿¥œã®ããŒã¯ã³ãäºæž¬ããŠçæããã
ãããã
ããŒã£ãšã
ãã®ãããéãã§ç¹°ãè¿ãããšã§ã
ãªã¢ã«ã¿ã€ã ã§éåããªã鳿¥œãçãŸããã£ãŠããã
éå»ã®10ç§éã®æèãèªãã§ã
次ã®2ç§ã®æç« ãæžãç¶ããã
ã¿ãããªã€ã¡ãŒãžããªã
è³¢ããªãã
ããããã®ããŒã³ã¿ ãªã¢ã«ã¿ã€ã ã
æ§èœãããããã ãã
MusicGenãšãã
Stable Audio Openã¿ãããªã
ä»ã®æåãªé³æ¥œçæã¢ãã«ãšæ¯ã¹ãŠãã
ãã©ã¡ãŒã¿æ°ã
ã€ãŸãã¢ãã«ã®èŠæš¡ã¯å°ããã®ã«ã
鳿¥œã®åè³ªãæž¬ãããã€ãã®ææšã§ã¯ã
ãããã®ã¢ãã«ãäžåã£ãŠããã ã
äŸãã°ã
FDopenl3 ã£ãŠããã¹ã³ã¢ã¯ã
72.14ã§ã
ä»ã®ã¢ãã«ãããã£ãšäœãã
ããã¯ã
çæãããé³ãã
ããèªç¶ã§æ¬ç©ã£ãœãã£ãŠããšã瀺ããŠããã ã
ããšã
é¢çœãå®éšããã£ãŠãã
ããã³ãã ãã©ã³ãžã·ã§ã³è©äŸ¡ã£ãŠããã®ã
ããã¯ã
äŸãã°ã
éããªã¢ã³ããšã³ããã¥ãŒãžãã¯ããã
æ¿ããããŒãããã¯ãžã
ã¿ããã«ã
äºã€ã®ç°ãªãã¹ã¿ã€ã«ã®éãã
ã©ãã ãã¹ã ãŒãºã«ç§»è¡ã§ãããããã¹ããããã ã
ãã®ã¢ãã«ã¯ã
60ç§ãããŠã
ã ãã ã鳿¥œã®ã¹ã¿ã€ã«ãå€åãããŠãããã ãã©ã
ãã®ç§»ãå€ãããã
ãããèªç¶ã§é³æ¥œçãªãã ã£ãŠã
éäžã§æ¥ã«ããã£ãšåãããããªãã§ã
åã®ã¹ã¿ã€ã«ã®èŠçŽ ãæ®ãã€ã€ã
æ°ããã¹ã¿ã€ã«ã«æº¶ã蟌ãã§ããã
ãŸãã«ã
ã©ã€ãDJãæ²ãããã¯ã¹ããŠãã¿ããã ããã
ãããã
ãã®ã©ã€ã ãã¥ãŒãžã㯠ã¢ãã«ã
ãªã¬ãã¡ã®ç掻ã®äžã§ã
ã©ããªé¢šã«äœ¿ãããã ããã
å¿çšäŸãèããŠã¿ãããã
ãŸãäžã€ç®ã¯ã
ãã£ã±ã鳿¥œã®ã©ã€ãããã©ãŒãã³ã¹ã ãã
DJãã
ãã®å Žã®é°å²æ°ãèŠãŠã
ããã¹ãããã³ãããæã¡èŸŒãã ãã§ã
æ²ã®ãžã£ã³ã«ã楜åšããªã¢ã«ã¿ã€ã ã§èªç±èªåšã«å€ããããã
ãã¯ããããžã£ãºãžã
ãããŠã¯ã©ã·ãã¯ãžã
ãªããŠããã
å代æªèã®ããã¯ã¹ãã
ã·ãŒã ã¬ã¹ã«å®çŸã§ãã¡ãããããããªãã
äºã€ç®ã¯ã
ã²ãŒã ã®BGMã ãã
ä»ã®ã²ãŒã ãã
å Žé¢ã«åãããŠBGMãå€ãããã©ã
åºæ¬çã«ã¯ã
ãããããçšæãããæ²ãã«ãŒãåçããŠãããã
ã§ããã®æè¡ã䜿ãã°ã
ãã¬ã€ã€ãŒã®è¡åãææ
ã«åãããŠã
BGMãç¡éã«ã
ãããŠãªã¢ã«ã¿ã€ã ã«çæãããããã«ãªãã
æ¿ããæŠéã·ãŒã³ã§ã¯ã
ãã¬ã€ã€ãŒã®æ»æã«åãããŠãã©ã ãæ¿ãããªã£ããã
åºå€§ãªãã£ãŒã«ããæ¢çŽ¢ããŠãããšãã¯ã
æ¯è²ãæéã«åãããŠã
ç©ããã§ã
ã§ã決ããŠåãã§ã¯ãªãæ²ãæµãç¶ãããã
ã²ãŒã ãžã®æ²¡å
¥æãã
ãšãã§ããªãããšã«ãªãããã ããã
äžã€ç®ã¯ã
ãã£ãšèº«è¿ãªé³æ¥œå¶äœãã
楜åšã®ç·Žç¿ã®ããŒãããŒãšããŠã
ãã®è«æã«ã¯ã
ãªãŒãã£ãª ã€ã³ãžã§ã¯ã·ã§ã³ã£ãŠããã
ãããé¢çœãæ©èœã玹ä»ãããŠããã ã
ããã¯ã
ãŠãŒã¶ãŒããã€ã¯ã«åãã£ãŠæã£ããã
ã®ã¿ãŒã匟ãããããé³ãã
AIããªã¢ã«ã¿ã€ã ã§åã蟌ãã§ã
ãã®é³ã«åãããç¶ãã®é³æ¥œãçæããŠãããæ©èœãªãã ã
ã€ãŸãã
ãªã¬ãã¡ãã£ãšéŒ»æãæããšã
AIãããã«åãããŠã
ãã£ãããããŒã¹ã©ã€ã³ãšãã©ã ãä»ããŠãããã
ã¿ãããªããšãã§ããã
楜åšãå§ããã°ããã®äººãã
ç°¡åãªã¡ããã£ã匟ãã ãã§ã
ãŸãã§ãã«ãã³ããšäžç·ã«ã»ãã·ã§ã³ããŠãã¿ãããªäœéšãã§ãããã ã
ããã
ç·Žç¿ããã¡ããã¡ã楜ãããªããšæããªãã
ä»ã«ãã
çŸè¡é€šã®ã€ã³ã¹ã¿ã¬ãŒã·ã§ã³ã§ã
芳客ã®åãã話ã声ã«åå¿ããŠã
空éå
šäœã®é³æ¥œãåžžã«å€åãç¶ããã¢ãŒãäœåãšãã
è²ã
ãªå¯èœæ§ãèããããããã
ãããŒã
鳿¥œãäœãã£ãŠããè¡çºã®æŠå¿µãã
æ ¹æ¬ããå€ãã£ã¡ãããããããªãã
ãããªå¯èœæ§ãæããããŠãããã
ãããè«æã ã£ããªã
åãªãéå
·ãããªããŠã
äžç·ã«åµé ããããŒãããŒãšããŠã®AIã
ãããªæªæ¥ãã
ãããããããŸã§æ¥ãŠãã®ãããããªããã
ãµã
ã
仿¥ã¯ãããªãšããããªã
ãŸãé¢çœãè«æèŠã€ãããã
玹ä»ãããã
ãããããã
ãŸãããŒã
äºã®å
ãã£ãä»®ã§ããã
ð The Paper and Some Imagination (English)
ð
ð TitleïŒAI Music Goes LIVE: Google DeepMind's Real-Time Breakthrough
ð Summary (English)
Hello everyone! It's 2025, August 8th, a Friday!
Today, I'm digging into the archives to pull out a real banger of a paper,
and this one, oh boy, itâs a game-changer for anyone who loves music.
Letâs get into it.
The title is
Live Music Models.
The URL is
https://arxiv.org/abs/2508.04651v1
It's long!
Yeah, they really need to work on making these URLs a bit snappier.
Alright, so whatâs this all about?
The folks at Google DeepMind are tackling a really cool idea.
They say music exists in two ways.
First, there's music as a noun.
You know, a recorded song, an MP3 file, something static you can just listen to.
Thatâs how most of us consume music, and it's how most AI music generators work.
You give them a prompt, wait a bit, and they spit out a finished track.
But then, there's the second form, music as a verb.
This is the live stuff, the performance, the act of creating music in the moment.
It's about that feeling of creative flow, that connection you get at a concert.
And that's the part that AI has really struggled with.
Until now, maybe.
The problem this paper is trying to solve is that gap.
Current AI music models are offline tools.
Theyâre not interactive instruments you can jam with in real-time.
Thereâs always a delay, a waiting period.
You can't have a back-and-forth musical conversation with them.
They're not built for the spontaneity of a live performance.
So, this paper introduces a new kind of AI,
something they call a live music model.
And they've released two of them, Magenta RealTime and Lyria RealTime.
The whole point of these models is to generate a continuous stream of music,
in real-time, that you can control as it's happening.
No more typing a prompt and waiting.
You're in a constant loop of creating and listening,
which is, you know, what playing an instrument is actually like.
What makes these models "live" are three key things.
One, they generate music faster than it takes to listen to it.
Two, they generate it as a continuous stream,
always building on what just came before it.
And three, the controls are super responsive with very low delay,
so you can actually interact with it meaningfully.
And you know what's really impressive?
Their open-weights model, Magenta RealTime,
actually performs better on some quality tests than bigger models,
like MusicGen and Stable Audio Open.
It uses 38 percent fewer parameters than Stable Audio Open,
and a whopping 77 percent fewer than MusicGen.
Thatâs seriously efficient.
Okay, so how does it work under the hood?
Itâs pretty clever, actually.
They use something called a codec language model.
Basically, they take a piece of audio,
and an AI called a codec squishes it down into a sequence of discrete tokens.
Think of it like turning a soundwave into a sentence made of special "audio words".
Then, a language model, which is great at predicting the next word in a sentence,
is trained to predict the next "audio word" in the sequence.
To make it work live, it operates in two-second chunks.
The model is always listening to the last ten seconds of music,
and based on that history, plus your instructions,
it generates the next two seconds of audio.
This chunk-based approach means it can just keep going forever,
creating an endless stream of music.
Right, so where could we actually use this?
This is the fun part. The applications are pretty wild.
First, for music production and practice.
Imagine you're a songwriter or a musician at home.
You could have an AI jam buddy that never gets tired.
You could start with a simple guitar riff,
and the AI could generate a bassline and drums that follow you.
If you want to change the mood, you just type,
"make it more funky" or "add a dreamy synth pad".
The AI would smoothly transition, creating a dynamic backing track on the fly.
It's like having a full band in your computer, ready to improvise with you.
Second, let's talk about video games.
This could totally revolutionize game soundtracks.
Right now, most game music is a set of pre-recorded loops that fade in and out.
It can feel a bit repetitive.
With a live music model, the music could react to everything the player does,
in real-time.
Imagine you're exploring a quiet forest, and the music is ambient and gentle.
Suddenly, a monster jumps out!
The AI could instantly shift the music to a high-energy battle theme,
with the tempo and intensity perfectly matching the action on screen.
Every player's experience would have its own unique, dynamic soundtrack.
That's a whole new level of immersion.
And third, for live performance and DJing.
This is where it gets really crazy.
A DJ could use text prompts to blend genres in completely new ways,
creating transitions that would be impossible with traditional turntables.
But there's an even cooler feature they call audio injection.
This audio injection thing is mind-blowing.
A performer could sing a melody, play a short drum pattern,
or a riff on a keyboard, and feed that audio directly into the model.
The AI doesn't just play it back.
It takes that audio as inspiration.
It might harmonize your melody,
or transform your drum pattern into a complex beat,
or weave your keyboard riff into the ongoing track with a different instrument.
Itâs true human-machine collaboration, live on stage.
The model can even blend styles using prompts.
You could give it a prompt for "techno" and a prompt for "flute",
and it would generate something that sounds like techno with a flute.
You can even give it an audio prompt,
like a sample of a song you like,
and it will try to match that style.
Of course, it's not perfect yet.
There's still a delay of about two seconds between your input and the AI's response.
And its memory is only ten seconds long,
so it canât remember a theme from the beginning of a song and bring it back later.
It can't create complex, long-form song structures on its own.
But the team knows this, and they're looking ahead.
They want to get the latency down to almost zero.
If they can do that, you could control it directly with a MIDI keyboard,
or even your voice, turning it into a whole new kind of synthesizer or effect.
The future they imagine is one where AI can be a true musical partner,
an improvising bandmate that listens and responds.
And honestly, that is an incredibly exciting future for music.
What a time to be alive.
ðïž ã³ã¡ã³ã
æåŸãŸã§èªãã§ãããŠæ¬åœã«ããããšãïŒïŒ
ãã€ãã©ãããããŸã話ããªããïŒããããããããããïŒ
Original paper link:ð
ãé¢é£ããŒã¯ãŒãã#AI #人工ç¥èœ #鳿¥œçæ #ãªã¢ã«ã¿ã€ã 鳿¥œ #ã©ã€ãããã©ãŒãã³ã¹ #ãã¥ãŒãžããã³ã° #è«æè§£èª¬ #ãµãŠã³ã #é³é¿ #MagentaRealTime #LiveMusicModels #arXiv #ãªãŒãã£ãªAI #ã²ãŒã BGM #楜åšç·Žç¿ #AIMusic #GoogleDeepMind #LiveMusicModels #RealTimeAI #MusicTech #GenerativeAI #AIMusicGeneration #MagentaRealTime #LyriaRealTime #AIForMusicians #MusicInnovation #TechBreakthrough #FutureOfMusic #InteractiveAI
