Glonce

YuE2 outperforms Suno v5 on WildSongBench music benchmark

modelsThe Glonce editorsHacker News🔥64Glonce rating64Impact55Novelty50Relevance80Surprise40Players25Credibility85

A project page titled 'YuE2: Frontier Music with Symbolic Planning' presents three separate music AI models alongside a set of interactive listening demos. YuE2 generates songs, MERT2 produces audio representations for tasks such as tagging and genre recognition, and SheetSage2 transcribes audio into musical elements such as beat, key, and chords. The page does not name who built or released the models: it refers to them only in the first person, as 'our models' and 'we', and gives no release date for the models themselves, no linked paper or code repository, and no license.

For YuE2, the page reports a best-of-8 score of 6.9632 on a metric called SongBench, in an evaluation dated September 5, 2026. That figure is described as the highest mean among 15 system settings tested on WildSongBench, a 192-prompt benchmark scored across seven metrics. Suno v5, evaluated in the same comparison, scores 6.8721. The page adds that rankings vary by metric, so the SongBench mean captures only one slice of the comparison; it does not state what scale the score is measured on, or give any statistical significance test for the gap between YuE2 and Suno v5. The best-of-8 result is not simply the top score out of eight generations either: the page says the chosen output is picked by musicality, prompt control, and lyric accuracy. A separate figure plots two composite axes, song quality (combining SongBench and SongEval) and text alignment (combining MuLan, AllMusicCaps, and prompt control), with bubble size showing AudioBox production quality.

MERT2 is offered as two encoder variants, MERT2-30s and MERT2-FS, each with 632 million parameters; MERT2-30s trains on a 30-second context window, while MERT2-FS trains on a 300-second, full-song window. Taking the better of the two encoders on each metric, the page reports state-of-the-art scores on 14 of 15 metrics in the MARBLE benchmark, against nine published baseline systems, leading across tagging, key, genre, and emotion recognition. Two examples given are a genre accuracy score of 91.72 on the GTZAN dataset, from MERT2-30s, and a key refined accuracy score of 67.05 on GiantSteps, from MERT2-FS; both are reported as the underlying score multiplied by 100. The page also states that MERT2's best-observed results use representations chosen according to test-set scores, while each individual ROC-AUC and average-precision pair is computed on one consistent representation.

SheetSage2 is presented as a single model covering six transcription tasks: beat, downbeat, key, chord, structure, and melody. The page reports state-of-the-art results on 10 of 13 benchmark metrics across these tasks, comparing SheetSage2 against SheetSage1, Madmom, and other task-specific systems, using one model selected by validation loss rather than a separate model per task. Cited examples include a pitch-class note F1 score of 82.51 for vocal melody transcription on RWC-Pop, and a chord recognition score of 90.08 using a major/minor vocabulary on osu2017. On a third dataset, Chords1217, the comparison is not run under identical conditions: the page states that the ChordFormer baseline uses five-fold cross-validation, while SheetSage2 is evaluated as a single fixed model across all 1,217 tracks.

On training data, the page says its models are trained primarily on CC0-licensed music and synthetic audio, with a company called Tokenwave.AI supplying most of the synthetic training data under license; it adds a general statement about being committed to the ethical and responsible use of data, without further detail. Training-data volumes differ sharply by model: 346K hours for YuE2, 700K hours for MERT2, and 28.4K hours for SheetSage2. Alongside the benchmark tables, the site offers several interactive demos: a 'Selected score' view that lets a visitor listen to a song and then explore its melody, rhythm, and chords as a symbolic plan; a 'Cover & Editing' tool that reshapes an existing song's melody, lyrics, tempo, or arrangement; an 'Agentic music editing' demo that edits a song through a conversational interface; and a 'Genre Explorer' that gathers listening examples across genres and languages. Full benchmark results for WildSongBench, MARBLE, and the SheetSage2 tasks are each available to download from the page.

Key facts

  • YuE2's best-of-8 output scores 6.9632 on SongBench, the highest mean among 15 system settings evaluated on the 192-prompt WildSongBench (evaluation dated September 5, 2026); Suno v5 scores 6.8721 in the same comparison, though the page notes that rankings vary by metric.
  • MERT2, built as two 632-million-parameter encoder variants with 30-second and 300-second (full-song) training contexts, reaches state-of-the-art on 14 of 15 MARBLE benchmark metrics against nine published baselines, including a genre accuracy score of 91.72 on GTZAN and a key refined accuracy score of 67.05 on GiantSteps.
  • SheetSage2, a single model spanning six transcription tasks (beat, downbeat, key, chord, structure, melody), reaches state-of-the-art on 10 of 13 benchmark metrics, including a pitch-class note F1 score of 82.51 on RWC-Pop and a chord recognition score of 90.08 on osu2017.
  • Training data is drawn primarily from CC0 music and synthetic audio, with Tokenwave.AI supplying most of the synthetic material under license; reported training volumes are 346K hours for YuE2, 700K hours for MERT2, and 28.4K hours for SheetSage2.
  • The site pairs the benchmarks with interactive demos, including a 'Selected score' view that breaks a song into its symbolic melody, rhythm, and chords, a 'Cover & Editing' tool that reshapes an existing song, and an 'Agentic music editing' demo that edits a song through conversation.

Why it matters

The page claims leading or state-of-the-art results across three different music AI tasks at once: generation (YuE2 against Suno v5), audio representation (MERT2 on the MARBLE benchmark), and transcription (SheetSage2 across six tasks). Together they cover more than song generation: MERT2 is presented as leading across tagging, key, genre, and emotion recognition, and SheetSage2 turns audio into structured musical data such as beat, chord, and melody. The generation result is also framed as a direct comparison to Suno v5, a commercial song generator named explicitly alongside YuE2's own score.

Who it affects

Musicians, producers, and hobbyists get two consumer-facing demos: a song generator and a conversational song-editing tool. Companies building competing song-generation products, Suno among them, are placed in a public benchmark comparison through the Suno v5 figure. Researchers and toolmakers working with music tagging, recognition, and transcription are the more direct audience for MERT2 and SheetSage2: MERT2's representations target tagging, key, genre, and emotion recognition, and SheetSage2 covers beat, downbeat, key, chord, structure, and melody transcription.

How to use it

Everything on the page is presented through the browser. The 'Selected score', 'Cover & Editing', 'Agentic music editing', and 'Genre Explorer' sections are interactive demos for listening to and manipulating songs, and the benchmark tables for WildSongBench, MARBLE, and the SheetSage2 tasks each carry a download link for the underlying results. The page does not say whether the three models themselves, their weights, or their code can be downloaded or run elsewhere, and it names no license for them, only that their training data is CC0 or synthetic.

How solid is it

The results are self-reported on the project's own page, with no linked paper, preprint, code repository, or technical report to check the methodology against. The core generation claim rests on a gap of 6.9632 against 6.8721, on a SongBench scale whose maximum value the page never states, so what either number is out of is not given; no statistical significance test or margin is given for the gap either. The page itself qualifies the claim by noting that rankings vary depending on which of the seven WildSongBench metrics is used, so the 'highest mean' result is one aggregate view rather than a win on every metric. For MERT2, the page states that its best-observed numbers use representations chosen according to test-set scores, while each individual ROC-AUC and average-precision pair keeps one representation fixed. For SheetSage2, the page discloses its own asymmetry on the Chords1217 dataset: the ChordFormer baseline is scored with five-fold cross-validation, while SheetSage2 is scored as a single fixed model on all 1,217 tracks, which the page itself distinguishes as a different test.

Risks and caveats

No individual, team, or organization is named anywhere on the page as the creator of YuE2, MERT2, or SheetSage2: the only voice is an unnamed 'we'. No release date is given for the models, only the September 5, 2026 date of the WildSongBench evaluation. No license is stated for the models or any code, and the page does not say whether they are open-sourced. No parameter count or architecture detail is given for YuE2 or SheetSage2 themselves; the only stated parameter count, 632 million, covers the MERT2 encoders alone. Training data is described only as 'primarily' CC0 music and synthetic audio, which leaves room for an unspecified remainder.