Excited to share SingSong, a system which can generate instrumental accompaniments to pair with input vocals!. 📄🔊. Work co-led by myself, @antoine_caillon, and @ada_rob as part of @GoogleMagenta and the broader MusicLM project 🧵

RT @noahconst: Made a walkthrough vid for Magenta RealTime “Audio Injection”! The notebook takes ~10m to spin up, but totally worth it for….

What do you want to know?.

RT @sedielem: In 2018, we wrote a paper titled "the challenge of realistic music generation". A few years ago, this challenge was essential….

🔊 Easy to lose hours to live jamming w/ Magenta RT. Feels like a glimpse at something magical. Try it yourself! Some setup involved (we're working on improving that!) but worth it:

RT @neuro_stack: very fun, i was just using the colab demo but i want to learn more about this

RT @hollyherndon: I’m happy to have donated my voice to this incredible new open model and instrument!. It is a new, active, approach to sp….

THIS DEMO 🔥🔥. My new fave metal-folk-techno track ft. @hollyherndon.

RT @Ilaria__Manco: We've worked with @hollyherndon and @matdryhurst to finetune Magenta RT on Holly's vocals and the result is a really co….

RT @osanseviero: Introducing. Perch 2! Our latest open model to analyze bioacoustic data 👀. 🐦Classify 15,000 species.🔊Generate audio embed….

You don’t need to watch the GPT5 livestream. The end game feature of live jamming with AI is already here!!.

RT @jesseengel: Realtime interactive generative models FTW! . Announcing a new 🌊 of details and features for Magenta RealTime, the open wei….

RT @lmarena_ai: 🧑‍🔬 Research Update: Today, we are releasing a new dataset with over 140k conversations from the text arena collected betwe….

RT @ml_angelopoulos: Check out Music Arena! . It really blows my mind how legitimately good today's music models are. With proper prompting….

RT @sethforsgren: Congrats on the launch @chrisdonahuey!. Excited for @riffusionai be a part of this, and for all these models to improve t….

Thanks to many others as well! @sonyai_global for financial support, @lmarena_ai for general guidance, and folks at @riffusionai @StabilityAI @GoogleDeepMind @SunoMusic @udiomusic for helpful feedback.

Music Arena is a work in progress! We will add new models and refine over time. It's crazy hard to build these platforms - we would never have gotten far without amazing collaborators @yonghyunk1m @iamwaynechi @ml_angelopoulos @infwinston @koichi__saito @shinjiw_at_cmu @mittu1204.

We're launching with a broad set of open and proprietary models from different providers. I am a bit scared about what will happen to our little 4 GPU machine serving the open weights models 😅, so please forgive downtime and let us know if you have spare compute to contribute 🙏

We aim to maximize transparency: the combination of our open source code and comprehensive data releases means anyone can audit the entire lifecycle of data on our platform, from prompts to audio to votes to (forthcoming) leaderboard.

Another challenge in music AI research is a lack of open and renewable preference data. We aim to fill this gap through a *rolling data release* policy: comprehensive monthly dumps of all the data we collect through the platform w/ user privacy protections.

While Music Arena follows evaluation trends in other AI areas, here we propose features specific to music. Music models have diverse type signatures - we use an LLM to route text prompts to appropriate ones. We also collect listening data for more fine-grained preference info

Human listening studies are the gold standard for evaluation in music AI, but study protocols are currently adhoc and unscalable. Music Arena aims to fill this gap through *live evaluation*: a standardized pairwise evaluation protocol that (hopefully!) scales w/ real-world users.