ACE Music v1 Ends Suno's Reign on Artificial Analysis's Blind Vocal Board
On this page 8 sections
TL;DR
ACE Music v1, a browser song generator from ACE Studio, debuted at number one on Artificial Analysis's AA-Music-Vocal v1.1 blind listening board: 1160 Elo against 1144 for Suno v6, with Mureka and Google's Lyria further back. The 16-point lead sits inside overlapping confidence intervals, and Suno still tops the separate instrumental board. The workflow is where ACE earns a look: performance directions inside your lyrics, MIDI and stems out the other end, and a preview version that writes an editable score before it sings. A lighter Flash model is promised open source in late 2026.
What the board says, precisely
Artificial Analysis runs its music ranking as a blind arena: evaluators hear two clips generated from the same prompt and pick the one they prefer, without knowing which model made which. Preferences become Elo, anchored so MiniMax Music 3.0 sits at exactly 1000. On the vocal board, ACE Music v1 leads at 1160 (±21) from 2,119 votes, with Suno v6 at 1144 (±18) from 3,063. ElevenLabs' Eleven Music v2.5 sits eighth at 1016.
Read the error bars before the headline. ACE's 95 percent interval runs 1139 to 1181 and Suno's runs 1126 to 1162, so the two ranges overlap and the top two are not statistically separated. Decrypt, which reviewed the model, makes the same point in plainer words.
The instrumental board is the other half of the story. Suno v6 leads AA-Music-Instrumental v1.1 at 1148, and ACE Music v1 has no entry there at all. One crown changed hands; the other stayed put.
Songs you can take apart
ACE Music v1 takes a text description or your own lyrics and returns a full song, which is now table stakes. Two choices separate it from the field. First, performance directions: put parenthetical cues before a line, like "(hum softly)" or "(sing with a raspy voice, add guitar)", and the model shapes delivery around them. You steer how a line is sung, not only what gets sung.
Second, the outputs are editable artifacts. A generation can be converted to MIDI and opened in your DAW, and a stem splitter isolates vocals, accompaniment or individual instruments for mixing. Decrypt's review also singles out the Idea feature, which develops a loose concept into lyrics plus a section plan you refine, instead of rerolling whole tracks until one sticks.
The v1.1 Preview goes further and plans the music as a score in ABC notation before generating any audio. You edit melodies and chords, then regenerate the performance. In the developers' words, "each generation is a new performance, so the sound may differ even when the score stays the same." Picture a session band that will replay the chart as many times as you like: change the sheet music, keep the musicians.
Who is ACE Studio
The company was founded in 2018 and started out in singing voice synthesis, turning MIDI and lyrics into vocals; its ACE Studio app is an AI music workstation. A vocals-first pedigree topping a vocals-first board is a tidy fit.
There is an open-source thread here too. ACE Studio and StepFun already ship ACE-Step 1.5, a separate model family with public code and weights that runs on your own hardware, below the flagship on quality. The official page also promises ACE Music v1 Flash, a local-friendly build balancing quality, speed and hardware requirements, open sourced in late 2026.
Hands-on: strong in the mainstream, lost in the metal
Decrypt's tests back the ranking for mainstream styles: more melodic and coherent songs, vocals with fewer synthetic artifacts, and cleaner endings, the place where Suno sometimes drifts into strange sounds, repetition or abrupt changes. Holding a musical idea together after the lyrics run out is the hard part, and ACE held it.
Niche genres are the failure mode. Asked for melodic death metal, ACE kept producing a pleasant clean female vocal, which is less Dimmu Borgir and more dimly lit wine bar. The review reports it confuses metal subgenres generally, while Suno kept the heavier arrangements and the stylistic consistency. Decrypt flags its own test as subjective rather than controlled, which is the right caveat.
The catches
- The lead over Suno v6 is 16 Elo with overlapping 95 percent intervals, on fewer votes (2,119 against 3,063). Treat first place as the reading of the week, not a settled fact.
- ACE Music v1 appears only on the vocal board. Suno v6 still leads instrumentals at 1148.
- The flagship is closed-source with no public API; generation happens on the official site only.
- Commercial rights require a paid plan, from $9.90 a month. Songs made on the free tier never gain commercial rights, and upgrading later does not backdate them.
- Flash going open source is dated only "late 2026". Until it ships, the open option is the weaker ACE-Step 1.5.
- Elo measures what blind listeners prefer on identical prompts. It says nothing about licensing, sample clearance or whether a track survives client review.
Why builders should care
If you build anything with generated audio, the artifact model here is the template worth copying: a score you can edit, a performance you can regenerate, MIDI and stems a DAW can actually use. That pipeline composes with real production tools, where a lone MP3 dead-ends.
If you make music for money, read the rights terms before the first render. Client work needs the paid plan from day one, and free-tier output stays non-commercial forever.
If you prefer local models, two dates matter: ACE-Step 1.5 runs today, and the stronger Flash is promised before the end of 2026. An open music model that ordinary GPUs can run is the piece this stack is still missing.
Key Takeaways
- ACE Music v1, from ACE Studio, debuted at number one on Artificial Analysis's AA-Music-Vocal v1.1 blind listening board with 1160 Elo, 16 points ahead of Suno v6 (1144), with Mureka V9.5 at 1101 and Google's Lyria 3.5 at 1050.
- The 95 percent intervals overlap (ACE 1139-1181, Suno 1126-1162), on fewer votes for ACE, and ACE has no entry on the instrumental board, where Suno v6 still leads at 1148.
- The distinguishing workflow: parenthetical performance directions inside lyrics, MIDI export, stem splitting, and a v1.1 Preview that plans music as an editable ABC-notation score before generating audio.
- The flagship is closed with no API; commercial rights require a paid plan from $9.90 a month, and free-tier songs never become commercial retroactively.
- ACE-Step 1.5, built with StepFun, is already open source and runs locally at lower quality; ACE Music v1 Flash is promised open source in late 2026.
- Decrypt's hands-on found stronger mainstream output and cleaner song endings, but Suno kept the edge on niche genres like melodic death metal.
Sources: Artificial Analysis, AA-Music-Vocal v1.1 leaderboard, Artificial Analysis, AA-Music-Instrumental v1.1 leaderboard, ACE Music v1 product page, ACE Music v1.1 Preview, ACE Music pricing, Decrypt, hands-on review (Nov 11, 2026), ACE-Step 1.5 on GitHub