Excited to share SingSong, a system which can generate instrumental accompaniments to pair with input vocals! 📄 https://t.co/1mRUaXvqVy 🔊 https://t.co/8RGezPu5YQ Work co-led by myself, @antoine_caillon, and @ada_rob as part of @GoogleMagenta and the broader MusicLM project 🧵

See https://t.co/Xj92707LZU to 🗳️ vote on your favorite, or to download a comprehensive data release for the preferences we've collected so far Thanks @yonghyunk1m @SonyAI_global @arena !

🎵Music Arena ⚔️ was accepted to the NeurIPS 2025 Creativity Track, and we've released a big update to celebrate! Includes new models from @SonautoAI and @elevenlabsio. Also, Music Arena is now available as a 🤗 @huggingface space and dataset!

For more information, we've prepared a 📰 blog post with all of these findings and more: https://t.co/k8SN2pShW0 ⚔️ Music Arena: https://t.co/1sTVjMfH09 📜 Paper: https://t.co/HqTg5PHxZm Data:

Thanks to blog post co-authors (@yonghyunk1m, Nathan Pruyne), all our collaborators (@iamwaynechi @ml_angelopoulos @infwinston @koichi__saito @shinjiw_at_cmu @mittu1204), and others (@sonyai_global, @lmarena_ai, @riffusionai_ )!

Mainnet countdown is here 🐙

We will be continuously updating Music Arena with new systems! Please reach out if you are interested in evaluating your music generation model on our platform

We collect natural language feedback in addition to binary prefs. Users tend to comment on both generation quality and prompt adherence. Sentiment analysis on this feedback is correlated with win rates, though also reveals new system-specific strengths and weaknesses!

Most of our users write their own prompts, as opposed to using one of our built in suggestions. User prompts emphasize genres, instruments, and modes, and most are very short (median length 7).

Ending EBT isn't about the money

We also observe a weak (ρ=0.082) but significant (p=0.012) correlation between the amount of time a user spends listening to a pair of outputs, and the overall "difficulty" of the comparison (codified by negative absolute difference in Arena score)

We collect nuanced listening behavior on Music Arena, revealing new insights E.g.: listening behavior differs dramatically between the 1st and 2nd tracks a user observes. Users listen to the 1st at length, then decide their preference after only the first few seconds of the 2nd

We aim for *comprehensive* transparency of the Music Arena platform. To this end, this first update comes paired with a comprehensive data release, to be updated on a rolling basis. ⚔️ https://t.co/00HvKj3zPx ⭐️ https://t.co/9f3mvsfco2

Music Arena went into public beta on July 28. In the first ~month of use, we collected 1051 votes on 1420 battles. We compile two leaderboards for instrumental-only (2/3 of votes, first tweet) and w/ vocals (1/3 of votes, below)

Imagine AI that cares about your health. WaterMinder’s AI Gulp Detection is like having a hydration coach in your pocket. Sip water → App detects it → Earn Rewards. Download now and see it work in real time.

Sharing our initial leaderboard and open data release for 🎶Music Arena⚔️! Music is subjective and multi-dimensional. A key goal of Music Arena is to provide insights beyond binary preferences! 🧵

Thanks to our team Yichen (Will) Huang, @zacknovack @Koichi__Saito @jiatongshi @_shinjiwatanabe @mittu1204 @jwthickstun and @SonyAI_global for the support! So don't be sad about your music gen evals, get MAD! 🎺😡🎺 https://t.co/2U2Oz4sHpS

To facilitate robust + reliable music gen eval, we release MAD as a drop-in replacement (w/MIT license!) for metrics like FAD/MMD, and MusicPrefs on HF for further eval research and preference modeling! https://t.co/2U2Oz4sHpS https://t.co/p9GJpLZbOl

We find that MAD has much stronger rank correlation (𝜏=0.62) with human preferences than FAD (𝜏=0.14)!

QM LIVE: @qmbigbeat and @cryptohondo with YOUR MORNING LOOK AT THE MARKETS

Second, we measure correlation between automatic metrics and MusicPrefs: a dataset of pairwise human prefs for text-to-music that we collect via MTurk. https://t.co/p9GJpLZJDT

We find that FAD lacks sensitivity to important desiderata like musicality. We propose an alternative based on our meta evaluation findings: FAD: *VGGish* embeddings + *Frechet Distance* divergence ⬇️ MAD: *MERT* embeddings + *MAUVE* divergence

We meta evaluate reference-based music evaluation metrics in two stages. First, we systematically search over a design space of embeddings and divergences (inclusive of FAD), using synthetic datasets capturing sensitivity to specific desiderata (e.g., musicality, diversity).

The capabilities of music generation models continue to improve, but progress on evaluation lags behind. Popular automatic metrics such as FAD were developed for different tasks (music enhancement) and their relevance to music gen is unclear. So how can we do better?

Concerned about biases and politics in your children's education? Be an advocate with these 6 simple steps. Follow THINC to learn more about what you can do as a parent.

Eval for music generation is notoriously ill-defined, but no fear! Presenting MAD, a new metric for music quality with stronger alignment to human preferences. Appearing at ISMIR this week! ⭐: https://t.co/2U2Oz4sHpS 📖: https://t.co/GxMDGmbSt1 🔊: https://t.co/gpw7OrEfz0 🧵