ringkernel-audio-fft

GPU-accelerated audio FFT processing with direct/ambience signal separation.

Overview

This crate demonstrates RingKernel's actor model by implementing a novel audio processing pipeline where each FFT frequency bin is an independent GPU actor. Actors communicate with neighbors via K2K (kernel-to-kernel) messaging to perform coherence analysis and separate direct sound from room ambience.

Architecture

Audio Input -> FFT -> [Bin Actors with K2K] -> Separation -> Mixer -> IFFT -> Output
                           |   |   |
                         neighbor links

Each frequency bin actor:

1. Receives its FFT data (magnitude + phase)
2. Exchanges data with left/right neighbors via K2K
3. Computes coherence metrics
4. Separates into direct and ambient components

Signal Separation Algorithm

The separation is based on inter-bin phase coherence:

• Direct signal: High phase coherence between neighboring bins (transient, localized sounds)
• Ambience: Low phase coherence, diffuse energy distribution (reverb, room tone)

Usage

use ringkernel_audio_fft::prelude::*;

#[tokio::main]
async fn main() -> Result<()> {
    let processor = AudioFftProcessor::builder()
        .fft_size(2048)
        .hop_size(512)
        .sample_rate(44100)
        .build()
        .await?;

    let input = AudioInput::from_file("input.wav")?;

    let output = processor
        .process(input)
        .dry_wet(0.5)       // 50% wet (ambience)
        .gain_boost(1.2)    // 20% gain boost
        .run()
        .await?;

    output.direct.write_to_file("direct.wav")?;
    output.ambience.write_to_file("ambience.wav")?;

    Ok(())
}

Components

Processing Parameters

Testing

cargo test -p ringkernel-audio-fft

The crate includes 32 tests covering FFT processing, coherence analysis, and signal separation.

License

Apache-2.0