Audio Generation

Whole songs, made on your Mac, from the chat. Ask the assistant for a song: it writes the lyrics and a style, and YuE2 — running on this machine through audio.cpp — plans a score and sings it. The song lands on the canvas as an audio clip.

What it does

The assistant writes sectioned lyrics and a style and hands them to Audio Generation. YuE2 first plans an editable score — melody, chords and sections, in ABC notation — and posts it in the chat, then renders the audio from exactly that score: 48 kHz stereo, up to about six minutes.

The song lands on an empty row at bar 1. Nothing of yours is trimmed or moved to make room for it.

Change it by asking: a new style, a new tempo, a reharmonised chorus, a melody-only cover over the same tune. Each version lands under the one it came from, which is muted, and a single ⌘Z takes the whole change back.

Using it

1. Open a chat and ask for a song — "write me a short indie folk song about the last train home".
2. The first time, a card asks before the model is downloaded (2.9 GB), and says what its license allows.
3. Watch the card: loading, writing the score, composing, rendering, decoding. Cancel stops it at once, and nothing lands.

Audio Generation: Song in the palette only explains; songs are made from the chat, because the lyrics and the style are what the assistant writes.

What it needs

• Your project, for writing — it inserts the song.
• The chat — the score, the progress and the receipt are posted there.
• Storage — it remembers what each song was made from, so an edit sends only what changes.
• One network download. On first use it fetches YuE2 3B (4-bit) from huggingface.co, 2.9 GB, once, against hashes carried in the signed catalogue. An interrupted download continues where it stopped. After that, every run is offline: nothing you write leaves your Mac.
• A Mac with Apple Silicon, macOS 13.3 or later, and 16 GB of memory or more. The model holds about 5 GB while it works and gives it all back when the song is done.

On an M2 Max a three-minute song takes about two and a half minutes at draft quality and six at final. Smaller chips take longer.

License

The engine, audio.cpp, is Apache-2.0. The YuE2 model weights are licensed CC BY-NC 4.0: what you make with them is for non-commercial use only. The download card says so before anything is fetched, and every song's receipt says it again.

Platforms

macOS, Apple Silicon only.