Engine (hush-core) · Hush AI docs

The speech enhancement engine of Hush AI, in real time on the CPU, with no Python, GPU or ONNX runtime. Two models:

NVIDIA Real-time RE-USE (Mamba), in plain Rust: Denoiser.
DeepFilterNet 3, with its official Rust inference (libDF 0.5.6 and tract) and the settings of its deep-filter command: DeepFilter. Output is bit-identical to libDF called directly.

Both implement Enhancer, the streaming interface that live::LiveWorker drives.

Stream a signal#

Rust

use hush_core::{Config, Denoiser};

let mut denoiser = Denoiser::with_config(
    "models/model.safetensors",
    Config { layers: 3, sample_rate: 16_000, look_ahead: 0 },
)?;
denoiser.set_automatic_worker_threads(
    hush_core::WorkerPlacement::Unrestricted,
    hush_core::WorkerScheduling::OsDefault,
)?;
let mut clean = Vec::new();
for chunk in noisy.chunks(480) {
    denoiser.process_into(chunk, &mut clean)?; // output as frames complete
}
clean.extend(denoiser.flush()?);

Sample rates: 8, 16, 22.05, 24, 32, 44.1 and 48 kHz. Layers: 3–12 (fewer is faster). Look-ahead: 0–2 frames (each adds 20 ms).

DeepFilterNet 3 runs at 48 kHz on one thread, in 10 ms frames, with 30 ms of delay:

Rust

use hush_core::{DeepFilter, Enhancer};

let mut model = DeepFilter::load("models/DeepFilterNet3_onnx.tar.gz")?;
let mut clean = Vec::new();
for chunk in noisy_48k.chunks(480) {
    model.process_into(chunk, &mut clean)?;
}
clean.extend(model.flush()?);

Live devices#

hush_core::live::LiveWorker connects audio callbacks to either model with lock-free queues, rate conversion, clock-drift control, meters and an A/B switch. Callbacks never allocate, lock or wait. hush_core::virtual_mic writes into the Hush AI Microphone drivers.

Examples#

Shell

cargo run --release -p hush-core --example enhance -- noisy.wav clean.wav 3
cargo run --release -p hush-core --example enhance_dir -- noisy48k/ out/ 0 0 models/DeepFilterNet3_onnx.tar.gz
cargo run --release -p hush-core --example bench_stream -- 16000 3 0 1050 50
cargo run --release -p hush-core --features devices --example audio_devices

Features#

Feature	Use
`devices`	Examples that open audio devices (`cpal`)
`wgpu`, `metal`, `vulkan`, `cuda`, `rocm`	The offline RE-USE model on a GPU (Burn)
`profile`, `trace`	Development diagnostics. Never measure latency with them

License#

Apache-2.0. The RE-USE weights are NVIDIA's, under the NVIDIA One-Way Noncommercial License. DeepFilterNet 3, libDF and tract are MIT or Apache-2.0.