Technology: the SKEIN Audio Compiler

SKEIN

Author Once. Compile Ahead of Time. Ship Anywhere.

Compiles For Any Target

Operating systems

WindowsmacOSLinuxEmbedded OSes

Silicon

NVIDIAAMDApple SiliconQualcommTINXP

Processors

CPUGPUNPU

API

C++CPython

The Libraries Underneath

100+real-time DSP and neural components behind the graph nodes

Demo

Hear a Compiled Graph

Three graphs, built and compiled in SKEIN, running against real audio. Pick a scenario, press play, flip the switches. Use headphones: clips are binaural, simulated from in-car measurements.

Running today on automotive SoCs from Qualcomm and TI, desktop GPUs from NVIDIA, AMD and Apple, and mobile NPUs

Why a Compiler

The Whole Graph,
Compiled in One Pass

SKEIN compiles the whole graph ahead of time into one real-time-safe binary, and the gains show up on the silicon customers already ship.

CUDA-class

FFT on par with hand-written CUDA in internal testing; 1.5–2× faster at small partitions

4×

faster than the silicon vendor's own FFT on a flagship mobile platform

Faster

than FFTW and PFFFT on Intel, Apple and Snapdragon CPUs

	GPU Audio SDK (runtime)	SKEIN (compiler)
Scheduling	On-device, while audio runs	Ahead of time, at build
Optimization scope	Per processor	Whole graph, compiled ahead of time
Targets	Ported and recompiled per GPU vendor	One graph, many targets — GPU, CPU and NPU
Performance	Best-effort	Known before you ship
Markets	Pro audio, DAWs, consumer GPUs	All of that, plus embedded, wearables, automotive, aviation

Both come out of pro audio — built for DAWs and a wide range of consumer GPUs — and SKEIN keeps all of that. What changes is when the work happens.

Foundation

The Layer Every Product
Runs On

GPUs were built for parallel processing millions of pixels every second. Audio is a sequential chain of interdependent steps, with almost no room for error. Our solution came from GPU ray tracing, and the same scheduling principles became G.E.A.R.S.

G.E.A.R.S.A dual-scheduler, lock‑free architecture coordinating thousands of GPU cores at the instruction‑set level with deterministic timing — sub‑millisecond latency for DSP and neural inference alikePatented 2024 (US 12,026,518 B2)

Platform MiddlewareThe libraries and runtime engine that let algorithms run across GPUs and NPUs — the layer every Soundscape and Soundcore solution ships on, in market since 2022

75×

processing acceleration vs CPU

2048+

concurrent audio threads

<1 ms

real-time neural network inference

5×

faster development via high-level APIs

SKEIN ships through PoC and commercial engagements, built with each client around their graph and hardware. SDK access on request.

Get in touch

One Graph. Every Target. Real-Time.

Author Once. Compile Ahead of Time. Ship Anywhere.

Compiles For Any Target

The Libraries Underneath

Hear a Compiled Graph

The Whole Graph,
Compiled in One Pass

CUDA-class

4×

Faster

The Layer Every Product
Runs On

75×

2048+

<1 ms

5×

Got an Idea? Let's Build It.

Soundscape

Soundcore

Company

Soundscape

Soundcore

Anything Else

One Graph. Every Target. Real-Time.

Author Once. Compile Ahead of Time. Ship Anywhere.

Compiles For Any Target

The Libraries Underneath

Hear a Compiled Graph

The Whole Graph, Compiled in One Pass

CUDA-class

4×

Faster

The Layer Every Product Runs On

75×

2048+

<1 ms

5×

Got an Idea? Let's Build It.

Soundscape

Soundcore

Company

The Whole Graph,
Compiled in One Pass

The Layer Every Product
Runs On