Auto Compose
Describe a sound, such as "a warm analog pad with a slow attack", and a language model sets up the patch. Then ask for changes, as in a conversation, until it sounds right.
Auto Compose is best at basic starting points. Complex patches still need adjusting by hand, and results vary widely from one language model to another.
Setting it up
Auto Compose is off until you set it up, and it needs a language model you provide: one running on your own computer, or a cloud service you have an account with.
- Choose Settings › Auto Compose and turn on Enable Auto Compose.
- Choose the API Format: OpenAI-compatible for LM Studio, Ollama, OpenAI and most other servers, or Anthropic Messages for Anthropic (Claude).
- Enter the Base URL: the server's address, such as http://127.0.0.1:1234 for LM Studio on this Mac. If using a cloud service, consult their documentation or perform an internet search for the API base URL.
- Enter an API Key if the service needs one. A local server usually does not.
- Click Fetch Models and choose a model, then click Test Connection. It reports how fast the model answers and whether it thinks before answering.
macOS allows plain http:// only to an IP address or a .local name. For a server with any other name, use https://.
A cloud service charges you for usage, usually based on which model you use. Each Send makes one or more requests, so keep close track of your usage and costs.
Composing
- Select a patch (a Temporary Area Part, to hear it as it changes) and choose the zone to compose.
- Click Auto Compose in the editor's header. A window opens for that patch and zone.
- For the first request, choose Start from: Init Patch to start clean, or This Patch to build on what is there.
- Type what you want and press Return (⇧-Return starts a new line).
The first time you send to a server, the app asks first and says what will be sent. The language model's reply says what it changed. The settings change in the editor immediately. If the model asked for something that could not be applied, the app asks it to correct itself once or twice, automatically.
- Each turn is one undo step: ⌘Z, in either window, takes it back.
- With Live Edit on for the Part, you hear each turn as it lands. Otherwise click Send to V-Synth when you are happy.
- Stop ends a turn that is taking too long.
- If the model changes something so the patch would be silent, such as a level at zero or a switch off, the app puts it right and says what it changed after "Made audible".
What it can change
The zone's oscillators (analog or PCM, with the wave), envelopes, LFOs, the Modulator, both COSM blocks, the Amplifier, the eight matrix controls, the patch's voice settings, and the multi-effects, chorus and reverb. It does not change key ranges, scale tuning, the Step Modulator, the arpeggio or the controllers.
What is sent
Each request carries what you typed, the conversation so far, the zone's settings, and the names of the waves it can use. It does not carry your notes, keywords or the library's name. The server's own terms and privacy policy apply to what it receives. See the privacy policy.
Other settings
| Setting | What it does |
|---|---|
| Let the model think before answering | Thinking takes longer. It mostly saves correction rounds rather than making better sounds. |
| Send full parameter list | Gives the model every effect and COSM type's settings at once. Turn it off for a small local model; changes of type then take two requests. |
| Tagging | Tags patches changed with the "Auto Compose assisted" keyword, and adds the first request and the model's name to the patch's note. |
| Log | Every Send is recorded on this Mac: rows applied and refused, timing and token counts. The API key is never written. Show in Finder opens it. |
Copy Prompt, in the Auto Compose window, puts the whole request on the clipboard: useful for seeing exactly what the model was told.