sf2subset

Trim a SoundFont 2 (.sf2) down to a chosen set of presets, keeping only the instruments and samples those presets use. Bank and program numbers are preserved, so code that loads a patch by program number needs no change.

Why

AVAudioUnitSampler (and most embedded synths) load .sf2 but not the Ogg-compressed .sf3, so a general-MIDI bank like GeneralUser GS ships as a ~31 MB file even when an app plays three patches. This tool cuts it to just those patches: for 24-TET Keys (piano/organ/synth) that is a 7.5 % file.

Polyphone can do the same by hand in its GUI, but this is a scriptable CLI so shrinking stays part of the build and survives a change of instrument set.

Install

go install github.com/botforge-pro/sf2subset@latest

This puts the sf2subset binary in $(go env GOPATH)/bin (usually ~/go/bin — make sure it is on your PATH). Pin a release with @v0.1.0 instead of @latest. For a one-off you don't even need to install: go run github.com/botforge-pro/sf2subset@latest -i in.sf2 -o out.sf2 -p 0,18,81. Or build from a clone with make build, which produces ./sf2subset.

Usage

sf2subset -i GeneralUser-GS.sf2 -o small.sf2 -p 0,18,81

• -i input .sf2
• -o output .sf2
• -p presets to keep: comma-separated bank:program (bank optional, defaults to 0), e.g. 0,18,81 or 0:0,0:18,0:81

It prints the kept preset/instrument/sample counts and the before/after size.

How it works

A SoundFont is a RIFF file with preset, instrument and sample tables that reference each other by index. sf2subset keeps the selected presets, walks their generators to the instruments they use, walks those to the samples they use (pulling in stereo partners via the sample links), then rebuilds every table with compacted indices and relocated sample data. Modulators and generators are copied verbatim, so the sound is unchanged.

Develop

make test   # go test ./...
make build  # go build -o sf2subset .
make check  # gofmt check + go vet + tests

The internal/sf2 package parses, writes and subsets; it is covered by round-trip and subset tests built on a small synthetic SoundFont, so the logic is verified without a large binary fixture.