sg-music-gen

An acestep.cpp-based command-line music generator for *nix.

1. Optionally, an NVIDIA GPU with a driver recent enough for the base image's CUDA; without one, generation runs on the CPU.

2. Install Git, Docker and make; for the GPU, also the NVIDIA Container Toolkit.

3. Clone this repository and cd to it:

$ git clone https://github.com/8vasu/sg-music-gen.git
$ cd sg-music-gen

1. Install:

$ make build && make models

2. Test, on the GPU if there is one, then on the CPU, then with the lyrics in lyrics.txt:

$ make test
$ make test-cpu
$ make test-lyrics

3. Clean up the test artifacts:

$ make clean

4. Remove the image and the downloaded models:

$ make remove

Symlink sg-music-gen.sh into a directory in your PATH, under any name, say sgsong:

$ ln -s "$PWD/sg-music-gen.sh" ~/.local/bin/sgsong

Symlink it rather than copying it: the script finds this directory by resolving the link. sgsong OUTPUT CAPTION writes OUTPUT.mp3, and OUTPUT.json with the lyrics and metadata the language model chose, into the current directory. Check the help message for all options:

$ sgsong -h

Examples:

$ sgsong rain "ambient piano, soft rain" --lyrics "[Instrumental]" --duration 60
$ sgsong summer "upbeat indie pop, female vocals" --duration 90
$ sgsong ballad "slow acoustic ballad" --lyrics "$(cat lyrics.txt)" --bpm 70 --keyscale "D major"
$ sgsong again "slow acoustic ballad" --seed 42 --output-format wav16
$ GGML_BACKEND=CPU sgsong cpu "lo-fi hip hop" --duration 30

On a GPU with little memory, add --vae-chunk 128; if the output is silence or noise on an older GPU, add --clamp-fp16.