This library can read SF2 SoundFont files and render audio samples from them in real-time. It properly reads in a compliant SF2 file and can be used to obtain meta data such as preset names. It also has an audio rendering engine that can generate audio samples for key events that come from (say) a MIDI keyboard. This library is currently being used by my SoundFonts application for SF2 file parsing, and more recently as an optional alternative to Apple's rendering offering.
Although most of the library code is generic C++17/23, there are a few bits that expect an Apple platform that has the AudioToolbox and Accelerate frameworks available. The goal is to be a simple library for reading SF2 files as well as a competent SF2 audio renderer whose output can be fed to any sort of audio processing chain, but it would probably take some effort to remove it from the Apple ecosystem.
Note that this package depends on some general DSP headers and audio classes from my AUv3Support package which is used by my various AUv3 extensions.
Currently, all SF2 generators and modulators are supported and/or implemented according to the SoundFont Spec v2. However, this library does not currently contain chorus or reverb effects. When rendering, the library will properly route a percentage of the signal to a chorus and reverb bus/channel if it is provided using the generator settiings that configure the percentage.
The render Engine
renderInto method takes a
Mixer instance which supports a main "dry" bus and two busses
for the "chorus effect send" and a "reverb effect send". These are populated with samples from active voices, and their
levels are controlled by the
reverbEffectSend parameters mentioned above. One can then connect
bus 1 to a chorus effect and bus 2 to reverb, and then connect those outputs and bus 0 of this library to a mixer to
generate the final output.
Here is a rough description of the top-level folders in SF2Lib:
There are quite a lot (yet not enough) unit tests that cover nearly all of the code base. There are even some rendering
tests that will play audio at the end if configured to do so. This option is found in the
Package.swift file, in the line
.define("PLAY_AUDIO", to: "0", .none). Change the "0" to "1" to
enable the audio output for all tests.
Alternatively, the unit tests with rendering capability have a
playAudio attribute which can
be set to
true to play the rendered output from a test. Note that the test results do not depend on this setting, but
enabling it does increase the time it takes to run the tests due to the time it takes to play back the recorded audio
The current code has not been optimized for speed, but it still performs very well on modern devices, including mobile. There are two tests (currently) that provide performance metrics:
As their name suggests, these exercise a specific interpolation method. They both generate 1 second of audio at a 48K sample rate, rendering 96 simultaneous notes. On an optimized build, the more expensive cubic 4th-order interpolation tests take ~0.27s to complete, or ~1/4 of the time budget. The faster linear interpolation is down to ~0.25s.
Addional performance gains could be had by following the approach of FluidSynth and render 64 samples at a time with no changes to most of the modulators and generators. Furthermore, one could check the pending MIDI event list to see if it is empty, and choose a path that supports vectorized rendering.
All of the code in this package is C++/Objective C but the Swift package now depends on Swift 5.9 in order to obtain
access to the Swift C++ bridging that is now available. This is only used in the
SF2Lib.cpp files to
declare a Swift-compatible interface to the C++ SF2::Render::Engine::Engine class, here named
interface creates and uses a pointer to an SF2::Render::Engine::Engine object so that it can hide all implementation
details from the Swift bridging framework.
All of the code has been written by myself over the course of several years, but I have benefitted from the existence of other projects, especially FluidSynth and their wealth of knowledge in all things SF2. In particular, if there is any confusion about what the SF2 spec means, I rely on their interpretation in code. That said, any misrepresentations of SF2 functionality are of my own doing.