Altinus

Thanks for all the support! Alt Denoiser is now live on GitHub. Windows VST3 binaries are available in the releases: github.com/Altinus/Alt-De…

Async inference with ONNX Runtime is now successfully loading the deep learning model. Since the model is relatively small, inference runs fine on CPU with reasonable resource consumption. No JUCE frameworks required at the bottom layer — it's pure Lua + a thin C++ wrapper

we can draw the spectrum waterfall in real time with new apis

Added more graphics apis to pdluajit, such as shadow and image support. Benchmarks show CPU usage drops dramatically from 18–23% to just 2–5% when repainting 8,000 shapes simultaneously.

After integrating luajit into pdlua, I find the syntax a little strange. For instance, C arrays created with ffi use 0-based indexing, while lua tables start at 1, which can definitely lead to confusion🤮

@Naskomusic I believe that, currently, it can load pure algorithm libraries focused on mathematical computations, rather than standard Pd externals which are bundled with Pd-specific API boilerplate? i haven`t try that though

@Altius__ wait wtf does this mean pdlua can load dll based externals??

The integration of JIT and FFI enhances lua`s DSP performance, achieving near-native C efficiency while enabling dynamic library invocation at runtime directly from Lua. dsp can be done simply by writing Lua scripts while maintaining very low CPU usage!

I’ve integrated LuaJIT into plugdata’s nightly build while keeping the original DSP API unchanged. Testing shows a performance boost for DSP tasks. For example, WSOLA processing went from ~13% CPU usage on Lua 5.4 down to just 2% with LuaJIT.

InkField | 墨域 is about to be released. From the first version, both humans and bots will be able to create on it together. In a way, the system reflects the time we live in. Humans and machines are already living in a kind of symbiosis. I’m generally an optimistic person. But I have to admit, 2026 feels a little heavy right now. Maybe the work will reflect that. Maybe it won’t. Either way, wishing everyone well. I still believe the dawn is coming.

Further tweaks, clones now smartly disable dsp when not needed, cutting down on resource usage.

@Naskomusic *oops, link is speechprocessingbook.aalto.fi/Representation…

@Altius__ ohhhhh i see!! does lowering it cause interference? i assume that‘s what the correlation is for right?

The real-time performance of pdlua is truly impressive. It has perfected the previous autotune and incorporated a new pitch-shifting algorithm.

@Naskomusic then accumulating the analysis increment approximates epoch positions; accumulating the synthesis increment triggers new grain emission

@Naskomusic I use improved YIN for pitch detection, basically an enhanced autocorrelation method. The resulting f0 is passed to psola to set grain length and the phase increments for analysis/synthesis windows. See: speechprocessingbook.aalto.fi/Representation…

@Naskomusic Pitching down formants is just slowing grain read speed, i uses linear interpolation here. This stretches each grain’s duration in the output, shifting formants lower. At moderate ratios. info loss is small i think, but artifacts do exist

@Altius__ When using PSOLA, are we losing some information when pitching down the formant? afaik we're remapping the content of the grain to expand, so naturally some information get's lost no? just going from intuition here i might be wrong!

Found a new approach that dramatically lowers the CPU load of my old plugdata granulator. On my pc, running 32 STEREO grains now takes around 7% CPU. By contrast, the old version with 32 MONO grains was eating up about 16%. demo(new one):

@Naskomusic ty!

@Altius__ that must be fun to use with a touchscreen :DD looks great btw!

just for fun: Implemented Karplus-Strong in pdlua, with a basic visualization UI added, with some AI assistance of course

@Naskomusic You can take a look at this code for reference: musicdsp.org/en/latest/Anal…. I’m calculating the autocorrelation using fft directly, so my approach differs somewhat from this. I’m not very familiar with the monophonic Helmholtz though

@Altius__ nice!! Are you doing auto correlation with an fft because it‘s easier compared to a monophonic helmholtz style one? i wouldn‘t even know where to begin with that in native pd 😅 but i assume it could also be done in pdlua?

vibe coded an LPC vocoder in pdlua💀 sped up autocorrelation with fft

@Naskomusic I use the autocorrelation + Levinson-Durbin algorithm to compute the LPC coefficients. The autocorrelation part is accelerated using rfft~, and the Levinson-Durbin recursion is implemented in pdlua