11. AI SOUND

AI SOUND lets you describe a sound in words; an AI then chooses the parameters and builds the patch. It can also make arpeggiator steps, chord progressions and MIDI phrases.

No AI is built into WAV3+CORE. You connect it to your own AI: a cloud AI used with your own API key, or a local LLM running on your computer (BYOK = Bring Your Own Key).

AI SOUND is a tab of the preset window: click the preset display (LCD) in the header and choose AI SOUND (or use AI Sound... in the LCD right-click menu, chapter 2). If you used the AI SOUND tab last time, the next click on the LCD opens it again. Esc, the × at the top right, or a click on the dark area outside closes the window. While the AI is working, a click outside or on another tab does not close it (use Stop, or Esc / × to stop and close).

• The text of your request.
• A description of WAV3+CORE and the list of parameters the AI may change (the range is set with PARAMS).
• The values of the current sound's parameters (only those that differ from the defaults).
• With Listen & refine on, the results of measuring the sound (numbers such as brightness and length; no audio is sent).

Depending on the provider, sent data may be stored or used to improve their services. Please check each provider's terms.

SAVE SETTINGS saves them.

• The API key is stored encrypted, readable only by your Windows user, in %APPDATA%\WAV3PLUSCORE. It is never written into presets or project files.
• The key is never sent over http:// (unencrypted) to anything other than your own computer (you get an error). Use https:// for cloud AIs.
• Never show your API key to anyone or paste it into chats or social media.

Screens, names, free-tier terms and model names of these services can change. This section reflects October 2026.

1. Open Google AI Studio (https://aistudio.google.com) in your browser and sign in with your Google account. Accept the terms the first time.
2. Click Get API key at the bottom left, then Create API key.
3. Copy the new key.

1. Click the preset display (LCD) and open the AI SOUND tab.
2. Choose Google Gemini under PROVIDER (the API URL is filled in).
3. Enter gemini-3.8-flash under MODEL. Adding a backup after a comma helps when the first is busy (e.g. gemini-3.8-flash, gemini-3.6-flash).
4. Paste the key under API KEY.
5. Click SAVE SETTINGS. PARAMS can stay at Auto (Full for cloud AIs).
6. Describe a sound in the box at the top and click GENERATE (e.g. bright supersaw trance lead).

About the free tier

• The free tier needs no card. Requests per minute and per day are limited; see the limits in AI Studio.
• Content sent on the free tier may be used to improve Google's products (per Google's pricing page). If that matters to you, use the paid tier (set up Billing) or a local LLM.
• If you get "503" or similar busy errors, wait a little or add a backup model in MODEL.

1. Prepare LM Studio

1. Download and install LM Studio from https://lmstudio.ai.
2. Find and download a model on the Discover screen. Start with a small one (e.g. Gemma 4 E4B = google/gemma-4-e4b) that fits in your computer's memory (VRAM).
3. Open the Developer screen (Local Server in older versions), load the model and start the server. Keep the default port 1234.
4. In the load settings, set Context Length (how much the AI can read at once) to about 16384. If it is too short you get "The AI ran out of room…".
5. For models with a thinking / reasoning feature, turning it off makes them more reliable.

1. Choose LM Studio (local) under PROVIDER (the API URL becomes http://localhost:1234/v1).
2. Enter the model identifier exactly as LM Studio shows it under MODEL (e.g. google/gemma-4-e4b). Copy and paste it to avoid typos.
3. Leave API KEY empty.
4. Click SAVE SETTINGS. Leaving PARAMS at Auto (Basic for local models) is recommended.
5. With the LM Studio server running, click GENERATE.

If it does not work

• "Set the model id.": MODEL is empty.
• "HTTP 400", "HTTP 404" and so on: check that MODEL matches the identifier shown in LM Studio, and that the model is loaded and the server is running.
• "The AI ran out of room…": increase Context Length.
• No reply / very slow: the model may be too large. Use a smaller model, or set PARAMS to Basic.

MIDI files made by the AI are saved in Documents\WAV3PLUSCORE\MIDI\AI. The MIDI FOLDER button (shown for Chord Progression / MIDI Pattern) opens the folder so you can drag the files into your DAW.

The AI is instructed not to copy existing songs.

• The AI draws a single-cycle wave by listing harmonic strengths. More harmonics give more detailed, brighter waves, but a longer reply that uses more tokens (cost).
• Up to 4 waves per reply, placed in Wave A-D of the oscillators (OSC1-6 when PARAMS is Full or above, otherwise OSC1-4).
• Each wave is stored, with a name, in an Add slot of ADDITIVE OSCILLATORS (chapter 9). Empty Add slots are used first; when none are free, a slot is overwritten.
• When a built-in wave is good enough, the AI picks it instead of drawing one.
• UNDO goes back to the sound before the AI changed it, drawn waves included.

• It can draw the LFOs whose shape PARAMS lets it choose: the OSC LFOs (OSC1-6 at Full and above, otherwise OSC1-4) and, from Standard, the Shape of Chorus / Flanger / Phaser / Hyper.
• The AI draws one cycle as points (position, height and the bend to the next point). The LFO it draws is set to Custom, and the shape is saved with the preset.
• When a built-in shape is enough, the AI picks it instead of drawing.
• If the points are unusable (too few, or flat with no movement), a sine-like shape is used.
• UNDO goes back to the sound before the AI changed it, drawn shapes included.

• For the measurement the arpeggiator and chord functions are switched off; C3 is held for 1.5 s, then a C major chord at C4 for 1.5 s.
• "Listened and checked: the AI is happy…" in the status line means the AI accepted the sound.
• It uses more requests, so it uses more tokens (cost).

More parameters make it easier for the AI to get lost, so Auto (Full) is recommended for everyday use. Choose Everything when Full is not enough.

Whatever the range:

• The MASTER volume is never raised above its current value. The FX LIMITER is always switched on (Ceiling −1 dB or lower). Separately, the last-stage safety limiter that protects ears and speakers (SAFE, chapter 1) is always active.
• When the AI does not set a mod matrix CURVE, it is chosen from the source (Uni-Linear + for envelopes, Velocity, Macros and so on; Bi-Linear + for LFOs and the like; chapter 6).
• A mod matrix slot the AI creates (in an empty slot, or by giving a slot a new source or destination) gets the same default Sync as ASSIGN (Step (All); chapter 6) unless the AI sets one. Slots where the AI only adjusts Range and the like, and slots it leaves alone, keep their Sync.
• Empty (reserved) choices and choices that load your own files (Conv Reverb's User File, Simulator's User IR) are never chosen.
• Every value the AI writes is checked first: numbers outside the range are clamped to the range, and unknown choice names or unreadable values are ignored (that control keeps its current value).

Tokens 21,505 in (88% cached) / 1,056 out · Session 85,038 / 1,402

Prices depend on the provider and model. Please check their price lists.

Troubleshooting

A log of the communication is kept in %APPDATA%\WAV3PLUSCORE\AI.log (plain text; the API key and anything after "?" in URLs are not written). It helps when something goes wrong.

Version 1.6.0 · 2026-11-08 · W3PC Audio · wav3pluscore.com