Choosing a model
Miso vendors eleven model specs from audio.cpp. Those eleven are what the Models screen lists and what the fourteen studio tasks run on.
audio.cpp supports more than this. Anything Miso did not vendor is dropped from the catalog, which is what keeps speech models off a music screen. If you have read the audio.cpp docs and are wondering where AudioSR, ControlFoley, Seed-VC or MiDashengLM went, that is why: they are not wired into Miso.
The matrix
Section titled “The matrix”| Family | Role | Vocals | Typical VRAM |
|---|---|---|---|
| ACE-Step 1.5 | Full songs, repaint, covers | Singing, 50+ languages | About 6.8 GB with memory saving |
| YuE2 | Full songs with an editable score | Singing, English and Chinese | About 9 GB peak in Q8 |
| MiniMax Music 3 | Full songs with structured lyrics | Expressive singing | 8 to 12 GB with memory saving |
| HeartMuLa | Tag-driven songs, long outputs | Multilingual singing | 8 to 10 GB with memory saving |
| Stable Audio 3 | Instrumentals and sound effects | No lyric singing | 6 to 8 GB small, about 12 GB medium |
| Mel-Band RoFormer | Vocal and instrumental split | Isolates vocals | 4 to 6 GB |
| BS-RoFormer | Vocal and instrumental split | Isolates vocals | 4 to 6 GB |
| HTDemucs | Four stem split | Isolates vocals | 3 to 4 GB |
| RVC | Voice conversion, four packaged voices | Converts singing | 3 to 4 GB |
| Vevo2 | Voice conversion, any voice, and singing | Converts and synthesises singing | 4 to 6 GB |
| MuScriptor | Audio to MIDI | Transcribes vocals and instruments | 4 to 6 GB |
Which song writer
Section titled “Which song writer”All four song writers do the same job differently. If you install one, install ACE-Step, because repaint and covers run on it too.
ACE-Step is the default and the one the guided prompt builder is designed around. It is also the only family with editing routes, so it is the only way to repaint or cover.
YuE2 is the only one that answers at 48 kHz stereo, and the only one that hands back a score you can edit and reuse. It works its own length out from your lyrics, so it has no duration control. It cannot write an instrumental: an empty lyric is refused outright. Its weights are CC-BY-NC-4.0, which is non-commercial.
MiniMax Music 3 wants a caption that reads like a production brief and lyrics with bracket tags. Strong dynamic range on the vocal. It is also the largest download at around 13 GB.
HeartMuLa takes a short description plus a list of comma-separated tags, and has a mode for long outputs.
Stable Audio 3 does not sing words at all. It is for instrumentals, and its other package set does sound effects.
Memory saving
Section titled “Memory saving”Most families have a mem_saver session option, and Miso sets it for you where the family supports it.
ACE-Step is the clearest case. Without memory saving its resident peak reaches 13.1 GB. With it, that drops to about 6.8 GB. On a card with 16 GB or less this is the difference between running and not.
One card, one model
Section titled “One card, one model”A language model and a 13 GB music model do not both fit on a single card. If you are running the lyrics assistant against a local llama.cpp server, free the card with Unload models before generating.
Models are almost all of the disk Miso uses. A full set runs past 40 GB. Your own projects are around 1 GB for a working set.
Settings has a Storage section showing both. See Storage and disk.