ACE-Step, 60 s
60.0 s of audio · inference 20.2 s · cold load 17.3 s · peak VRAM 44.4 GiB
Five audio models through the real Kollektive stage adapters on one disposable rented GPU: 155 recorded cases, 151 with audio. Listening copies are MP3 encodes of the verified WAV outputs. Every case shows its recorded outcome; quality_failure is yours to assign below.
| basket | completed | unsupported_runtime | unsupported_model | runtime_limit | model_limit | failed | skipped | quality_failure |
|---|---|---|---|---|---|---|---|---|
| ace | 22 | |||||||
| chatterbox | 51 | |||||||
| kokoro | 29 | |||||||
| kokoro-cpu | 10 | |||||||
| moss | 23 | 1 | 2 | |||||
| yue2 | 16 | 1 |
English, warm piano pop, expressive female voice, acoustic piano, rounded bass, light drums, lyrical memorable melody, unhurried phrasing, 88 BPM
[Verse] Streetlights fade along the lane Footsteps keep the time of rain Fold the night and leave it there Morning has a sky to spare [Chorus] Let the morning find us here Every window running clear Hold a little room for light We will sing beyond the night [Verse] Coffee steam and open doors Someone sweeping yesterday's floors Nothing here is only mine Every hand is in the line [Chorus] Let the morning find us here Every window running clear Hold a little room for light We will sing beyond the night
60.0 s of audio · inference 20.2 s · cold load 17.3 s · peak VRAM 44.4 GiB
114.1 s of audio · inference 88.1 s · cold load 18.5 s · peak VRAM 7.9 GiB
English, synthwave, confident male voice, analog synth arpeggios, gated reverb drums, driving bass, 110 BPM, nostalgic
[Verse] Tail lights drawing lines of red Radio says the storm's ahead Dashboard glow and empty seats Counting mile posts, counting beats [Chorus] Neon highway, take me far Past the last unbroken star I don't need to know the end Just the curve beyond the bend [Verse] Every city is a wave Every exit is a cave I keep driving, I keep still Motion is my only will [Chorus] Neon highway, take me far Past the last unbroken star I don't need to know the end Just the curve beyond the bend
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
197.2 s of audio · inference 112.1 s · warm load 0.0 s · peak VRAM 8.4 GiB
English, acoustic folk, intimate female voice, fingerpicked steel-string guitar, soft upright bass, sparse, 72 BPM
[Verse] There's a chip in the cup you gave me There's a crack where the light gets through And I pour the tea too early So it's cold by the time it's due [Chorus] Sing it low at the kitchen table Sing it slow so the walls can hear We were never rich or able But we never left from here [Verse] There's a map with a town we circled There's a coat that you never wore And I keep the door unlatched Like you're coming back for more [Chorus] Sing it low at the kitchen table Sing it slow so the walls can hear We were never rich or able But we never left from here
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
154.2 s of audio · inference 85.3 s · warm load 0.0 s · peak VRAM 8.3 GiB
English, boom bap hip hop, confident male rap, dusty drum break, upright bass sample, vinyl crackle, 92 BPM
[Verse] Started with a sketch on the back of a receipt Pinned it to the wall where the makers meet Nobody asked me who I knew They asked me what I'm gonna build from this, so I built it true Took a broken beat and a borrowed bass Left the credits in, kept the lineage in place [Chorus] Built from this, built from that Every line I write, I'm handing back No locked doors, no private key What you make from mine, you make for free [Verse] Queue is long but the queue is fair Put your name down, put your work out there Credits spent is a promise kept Every hour counted, every hour swept [Chorus] Built from this, built from that Every line I write, I'm handing back No locked doors, no private key What you make from mine, you make for free
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
76.6 s of audio · inference 42.3 s · warm load 0.0 s · peak VRAM 8.2 GiB
English, melodic metal, powerful male vocals, dual harmonised electric guitars, double kick drums, soaring anthemic chorus, 150 BPM
[Verse] Forged beneath a sleeping sky Every hammer asking why Sparks are falling like a rain Nothing tempered feels no pain [Chorus] Iron choir, raise your voice Every scar was once a choice Stand together, stand as one We are iron, we are the sun [Verse] They said the furnace would go cold They said the story had been told But the bellows breathe again And the anvil counts to ten [Chorus] Iron choir, raise your voice Every scar was once a choice Stand together, stand as one We are iron, we are the sun
60.0 s of audio · inference 2.0 s · warm load 0.0 s · peak VRAM 44.4 GiB
239.2 s of audio · inference 141.8 s · warm load 0.0 s · peak VRAM 8.6 GiB
English, jazz ballad, smoky female voice, brushed drums, upright bass, soft piano chords, muted trumpet, 62 BPM
[Verse] The barman's stacking chairs on tables The trumpet's packing up his case And I'm still holding half a story And the other half of your face [Chorus] Last call at the Blue Room One more for the road we never took Leave the light on in the doorway I'll be the shadow in the book [Verse] They'll sweep the ashes with the evening They'll pour the river down the drain But the song we didn't finish Is the one I hum in the rain [Chorus] Last call at the Blue Room One more for the road we never took Leave the light on in the doorway I'll be the shadow in the book
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
172.6 s of audio · inference 89.7 s · warm load 0.0 s · peak VRAM 8.3 GiB
Finnish, melancholic indie pop, soft male voice, clean electric guitar, warm synth pad, light drums, 96 BPM
[Verse] Katulamput syttyy yksitellen Lumi peittää jäljet eilisen Kannan kahvin kahdelle Vaikka tiedän, ettet tule [Chorus] Hiljainen katu, hiljainen talo Ikkunassa yksi valo Jos sä joskus kuljet tästä Jätän oven auki [Verse] Radiossa vanha laulu soi Sanat muistan, sävel unohtui Kaupunki hengittää hitaasti Ja minä sen mukana [Chorus] Hiljainen katu, hiljainen talo Ikkunassa yksi valo Jos sä joskus kuljet tästä Jätän oven auki
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
143.6 s of audio · inference 76.4 s · warm load 0.0 s · peak VRAM 8.3 GiB
English, upbeat funk, group vocals with call and response, punchy horn section, slap bass, wah guitar, tight drums, 112 BPM
[Verse] Got a Saturday engine and it won't sit still Got a pocket full of nothing and a window sill Turn the radio up till the neighbours knock We don't need a reason, we just need a clock [Chorus] Move it! (move it) to the left Shake it! (shake it) till there's nothing left Saturday engine, Saturday sun Everybody's working and it feels like fun [Verse] Bass line walking like it owns the street Horns are talking to the people's feet If you came to watch, then you came too late Get up off the curb, come and celebrate [Chorus] Move it! (move it) to the left Shake it! (shake it) till there's nothing left Saturday engine, Saturday sun Everybody's working and it feels like fun
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
91.3 s of audio · inference 54.1 s · warm load 0.0 s · peak VRAM 8.4 GiB
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
60.0 s of audio · inference 1.9 s · warm load 0.0 s · peak VRAM 44.4 GiB
Kokoro voices on the GPU and on the CPU; Chatterbox with its built-in voice and four public-domain reference readers.
Every valid Work enters Commons, and every Queue submission is a commitment. Please read both before you press Generate.
7.8 s of audio · inference 0.9 s · cold load 2.4 s
8.8 s of audio · inference 0.4 s · warm load 0.0 s
7.6 s of audio · inference 0.3 s · warm load 0.0 s
7.8 s of audio · inference 1.6 s · cold load 2.1 s
8.9 s of audio · inference 0.4 s · warm load 0.0 s
7.0 s of audio · inference 0.3 s · warm load 0.0 s
6.1 s of audio · inference 4.8 s · cold load 11.9 s · peak VRAM 3.1 GiB
7.6 s of audio · inference 9.5 s · warm load 0.0 s · peak VRAM 3.5 GiB
10.3 s of audio · inference 6.8 s · warm load 0.0 s · peak VRAM 3.8 GiB
6.8 s of audio · inference 4.7 s · warm load 0.0 s · peak VRAM 3.9 GiB
9.7 s of audio · inference 6.5 s · warm load 0.0 s · peak VRAM 3.9 GiB
Good morning, Helsinki. It is minus four degrees at seven fifteen, with light snow expected after two o'clock and winds of twelve metres per second.
10.2 s of audio · inference 0.4 s · warm load 0.0 s
10.9 s of audio · inference 0.4 s · warm load 0.0 s
9.6 s of audio · inference 0.4 s · warm load 0.0 s
10.2 s of audio · inference 2.2 s · warm load 0.0 s
11.2 s of audio · inference 6.9 s · warm load 0.0 s · peak VRAM 3.8 GiB
9.6 s of audio · inference 6.3 s · warm load 0.0 s · peak VRAM 3.9 GiB
11.2 s of audio · inference 7.2 s · warm load 0.0 s · peak VRAM 3.9 GiB
7.4 s of audio · inference 5.0 s · warm load 0.0 s · peak VRAM 3.9 GiB
9.8 s of audio · inference 6.6 s · warm load 0.0 s · peak VRAM 3.9 GiB
No, wait. I didn't mean it like that. Can we just start over, please?
4.3 s of audio · inference 0.2 s · warm load 0.0 s
4.9 s of audio · inference 0.2 s · warm load 0.0 s
4.4 s of audio · inference 0.2 s · warm load 0.0 s
4.3 s of audio · inference 0.7 s · warm load 0.0 s
6.2 s of audio · inference 4.1 s · warm load 0.0 s · peak VRAM 3.7 GiB
4.0 s of audio · inference 3.2 s · warm load 0.0 s · peak VRAM 3.9 GiB
6.2 s of audio · inference 4.4 s · warm load 0.0 s · peak VRAM 3.9 GiB
3.1 s of audio · inference 2.6 s · warm load 0.0 s · peak VRAM 3.9 GiB
4.2 s of audio · inference 3.4 s · warm load 0.0 s · peak VRAM 3.9 GiB
She sells seashells by the seashore, and the shells she sells are surely seashells.
5.0 s of audio · inference 0.2 s · warm load 0.0 s
5.6 s of audio · inference 0.3 s · warm load 0.0 s
5.2 s of audio · inference 0.3 s · warm load 0.0 s
5.0 s of audio · inference 0.8 s · warm load 0.0 s
8.2 s of audio · inference 5.2 s · warm load 0.0 s · peak VRAM 3.8 GiB
5.5 s of audio · inference 4.1 s · warm load 0.0 s · peak VRAM 3.9 GiB
8.2 s of audio · inference 5.4 s · warm load 0.0 s · peak VRAM 3.9 GiB
4.5 s of audio · inference 3.4 s · warm load 0.0 s · peak VRAM 3.9 GiB
6.6 s of audio · inference 4.7 s · warm load 0.0 s · peak VRAM 3.9 GiB
On the 3rd of March 2026, Dr. Aino Virtanen presented 1,284 measurements to the committee in under forty minutes.
9.5 s of audio · inference 0.4 s · warm load 0.0 s
10.2 s of audio · inference 0.4 s · warm load 0.0 s
9.1 s of audio · inference 0.4 s · warm load 0.0 s
9.5 s of audio · inference 1.7 s · warm load 0.0 s
9.9 s of audio · inference 6.1 s · warm load 0.0 s · peak VRAM 3.8 GiB
8.1 s of audio · inference 5.6 s · warm load 0.0 s · peak VRAM 3.9 GiB
9.9 s of audio · inference 6.4 s · warm load 0.0 s · peak VRAM 3.9 GiB
7.8 s of audio · inference 5.2 s · warm load 0.0 s · peak VRAM 3.9 GiB
9.1 s of audio · inference 6.2 s · warm load 0.0 s · peak VRAM 3.9 GiB
Have you ever wondered what a ten-thousand-year-old oak would say if it could speak?
4.8 s of audio · inference 0.2 s · warm load 0.0 s
5.3 s of audio · inference 0.3 s · warm load 0.0 s
4.7 s of audio · inference 0.2 s · warm load 0.0 s
4.8 s of audio · inference 0.8 s · warm load 0.0 s
5.5 s of audio · inference 3.6 s · warm load 0.0 s · peak VRAM 3.7 GiB
4.4 s of audio · inference 3.5 s · warm load 0.0 s · peak VRAM 3.9 GiB
5.5 s of audio · inference 3.9 s · warm load 0.0 s · peak VRAM 3.9 GiB
3.7 s of audio · inference 2.9 s · warm load 0.0 s · peak VRAM 3.9 GiB
4.8 s of audio · inference 3.8 s · warm load 0.0 s · peak VRAM 3.9 GiB
Breathe in slowly. Hold it for a moment. Now let it go, and notice how the room feels quieter than before.
6.9 s of audio · inference 0.3 s · warm load 0.0 s
7.7 s of audio · inference 0.3 s · warm load 0.0 s
6.8 s of audio · inference 0.3 s · warm load 0.0 s
6.9 s of audio · inference 1.1 s · warm load 0.0 s
9.7 s of audio · inference 6.1 s · warm load 0.0 s · peak VRAM 3.8 GiB
6.6 s of audio · inference 4.7 s · warm load 0.0 s · peak VRAM 3.9 GiB
9.7 s of audio · inference 6.3 s · warm load 0.0 s · peak VRAM 3.9 GiB
7.0 s of audio · inference 4.8 s · warm load 0.0 s · peak VRAM 3.9 GiB
9.1 s of audio · inference 6.1 s · warm load 0.0 s · peak VRAM 3.9 GiB
The river remembers every stone it carried, long after the mountain has forgotten them.
5.2 s of audio · inference 0.3 s · warm load 0.0 s
6.0 s of audio · inference 0.3 s · warm load 0.0 s
5.3 s of audio · inference 0.3 s · warm load 0.0 s
5.2 s of audio · inference 0.9 s · warm load 0.0 s
6.4 s of audio · inference 4.2 s · warm load 0.0 s · peak VRAM 3.7 GiB
5.4 s of audio · inference 4.2 s · warm load 0.0 s · peak VRAM 3.9 GiB
6.4 s of audio · inference 4.5 s · warm load 0.0 s · peak VRAM 3.9 GiB
4.4 s of audio · inference 3.3 s · warm load 0.0 s · peak VRAM 3.9 GiB
8.2 s of audio · inference 5.7 s · warm load 0.0 s · peak VRAM 3.9 GiB
10.0 s of audio · inference 33.8 s · cold load 28.6 s · peak VRAM 14.4 GiB
10.0 s of audio · inference 14.7 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.7 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
10.0 s of audio · inference 14.8 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
The checkpoint denoises one fixed 30-second latent and crops it; longer native generation is not possible without stitching, which is out of scope.
The admission lists transform and continue, but sound.generate.moss.v1 rejects Source audio until the pinned pipeline's conditioning path is wired.
The checkpoint denoises one fixed 30-second latent and crops it; longer native generation is not possible without stitching, which is out of scope.
Length ladders, planner modes, step counts, longer ambience and expressive extremes. Escalation stops at the first failed step; non-attempted cases carry their classification.
120.0 s of audio · inference 3.0 s · warm load 0.0 s · peak VRAM 44.4 GiB
120.0 s of audio · inference 3.0 s · warm load 0.0 s · peak VRAM 44.4 GiB
180.0 s of audio · inference 4.0 s · warm load 0.0 s · peak VRAM 44.4 GiB
180.0 s of audio · inference 4.0 s · warm load 0.0 s · peak VRAM 44.4 GiB
240.0 s of audio · inference 5.1 s · warm load 0.0 s · peak VRAM 44.4 GiB
240.0 s of audio · inference 5.1 s · warm load 0.0 s · peak VRAM 44.4 GiB
60.0 s of audio · inference 1.7 s · warm load 0.0 s · peak VRAM 44.4 GiB
4.4 s of audio · inference 3.4 s · warm load 0.0 s · peak VRAM 3.9 GiB
3.8 s of audio · inference 3.1 s · warm load 0.0 s · peak VRAM 3.9 GiB
6.4 s of audio · inference 4.5 s · warm load 0.0 s · peak VRAM 3.9 GiB
4.8 s of audio · inference 3.8 s · warm load 0.0 s · peak VRAM 3.9 GiB
5.6 s of audio · inference 4.1 s · warm load 0.0 s · peak VRAM 3.9 GiB
5.0 s of audio · inference 3.8 s · warm load 0.0 s · peak VRAM 3.9 GiB
6.6 s of audio · inference 4.8 s · warm load 0.0 s · peak VRAM 3.9 GiB
7.4 s of audio · inference 5.2 s · warm load 0.0 s · peak VRAM 3.9 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
30.0 s of audio · inference 15.9 s · warm load 0.0 s · peak VRAM 14.5 GiB
The checkpoint denoises one fixed 30-second latent and crops it; longer native generation is not possible without stitching, which is out of scope.
The checkpoint denoises one fixed 30-second latent and crops it; longer native generation is not possible without stitching, which is out of scope.
194.0 s of audio · inference 101.4 s · warm load 0.0 s · peak VRAM 8.5 GiB
271.1 s of audio · inference 163.8 s · warm load 0.0 s · peak VRAM 8.9 GiB
96.4 s of audio · inference 50.0 s · warm load 0.0 s · peak VRAM 8.3 GiB
134.5 s of audio · inference 59.8 s · warm load 0.0 s · peak VRAM 9.2 GiB
165.0 s of audio · inference 84.7 s · warm load 0.0 s · peak VRAM 8.4 GiB
137.3 s of audio · inference 61.4 s · warm load 0.0 s · peak VRAM 9.2 GiB
59.0 s of audio · inference 34.0 s · warm load 0.0 s · peak VRAM 8.3 GiB
93.4 s of audio · inference 53.4 s · warm load 0.0 s · peak VRAM 8.4 GiB
What a member could build from an existing Work: continuation, cover, repaint and variation through the real ACE handlers on this session's own verified outputs; unsupported paths are recorded, not forced.
60.0 s of audio · warm load 0.0 s · peak VRAM 44.4 GiB
60.0 s of audio · inference 2.4 s · warm load 0.0 s · peak VRAM 44.4 GiB
60.0 s of audio · inference 2.4 s · warm load 0.0 s · peak VRAM 44.4 GiB
60.0 s of audio · inference 2.4 s · warm load 0.0 s · peak VRAM 44.4 GiB
60.0 s of audio · inference 2.4 s · warm load 0.0 s · peak VRAM 44.4 GiB
The admission lists transform and continue, but sound.generate.moss.v1 rejects Source audio until the pinned pipeline's conditioning path is wired.
YuE2 audio covers go through SheetSage2 transcription into an ABC score; neither the transcriber nor the score input is in the admitted runtime or stage contract.
A 30-second paragraph and a two-minute narration. Chatterbox's single-call limit is about 40 seconds of speech; the over-limit case documents what happens, not a failure of the model.
27.2 s of audio · inference 16.5 s · warm load 0.0 s · peak VRAM 4.3 GiB
19.4 s of audio · inference 12.1 s · warm load 0.0 s · peak VRAM 4.1 GiB
40.0 s of audio · inference 24.3 s · warm load 0.0 s · peak VRAM 4.9 GiB
20.1 s of audio · inference 0.8 s · warm load 0.0 s
73.0 s of audio · inference 2.8 s · warm load 0.0 s
69.0 s of audio · inference 2.6 s · warm load 0.0 s
20.1 s of audio · inference 5.0 s · warm load 0.0 s
73.0 s of audio · inference 17.7 s · warm load 0.0 s
A40 at a quoted $0.567/h; rental 2026-10-18 10:30 UTC to removing_key. Nominal figures from measured inference time at the quoted rate; setup, downloads and idle time are excluded, and the provider's invoice is the only billing truth.
| basket | completed | measured inference | audio produced | nominal GPU cost | per minute of audio |
|---|---|---|---|---|---|
| ace | 22 | 73 s | 2040 s | $0.011 | $0.0003 |
| chatterbox | 51 | 288 s | 419 s | $0.045 | $0.0065 |
| kokoro | 29 | 15 s | 344 s | $0.002 | $0.0004 |
| kokoro-cpu | 10 | 32 s | 147 s | $0.005 | $0.0021 |
| moss | 23 | 365 s | 330 s | $0.057 | $0.0104 |
| yue2 | 16 | 1298 s | 2339 s | $0.204 | $0.0052 |