Sonora

Pure Rust implementation of WebRTC audio processing, providing echo cancellation, noise suppression, and automatic gain control.

Ported from the WebRTC Native Code (M145) audio processing module.

Crate	Description	Crates.io	Documentation
sonora	Full audio processing pipeline
sonora-ffi	C API (FFI) for integration with C/C++ projects
sonora-aec3	Echo Canceller 3 (AEC3)
sonora-agc2	Automatic Gain Control with RNN VAD
sonora-ns	Noise Suppression
sonora-common-audio	Audio DSP primitives (resamplers, filters)
sonora-simd	SIMD operations (SSE2, AVX2, NEON)
sonora-fft	FFT implementations (Ooura, PFFFT)

• Echo Cancellation (AEC3) -- adaptive filter-based echo canceller with delay estimation
• Noise Suppression -- Wiener filter-based noise reduction with voice activity detection
• Automatic Gain Control (AGC2) -- RNN VAD-based gain controller with limiter
• High-Pass Filter -- DC offset removal
• C API -- cbindgen-generated C header for FFI integration (via sonora-ffi)

Run the minimal echo cancellation demo:

cargo run -p sonora --example simple

More examples in crates/sonora/examples/:

The karaoke and recording examples require the examples feature which pulls in cpal, hound, and other audio I/O dependencies. These examples are based on the tonarino/webrtc-audio-processing examples, ported from PortAudio to cpal.

Runtime feature detection is used for AVX2 on x86_64. SSE2 is assumed available on all x86_64 targets. NEON is used on AArch64. A scalar fallback is provided for all other architectures.

The C++ reference test suite (WebRTC M145, 2400+ tests) is validated on Ubuntu x86_64 with the Rust backend linked via the sonora-sys FFI bridge.

Full pipeline processing a 10 ms frame with AEC3 + noise suppression + AGC2 enabled. Measured on Apple M4 Max (NEON backend), Rust 1.85, -C target-cpu=native:

See BENCHMARKS.md for per-component comparisons, profiling breakdown, and instructions for reproducing.

This project uses cargo-make for task automation.

cargo install cargo-make

cargo make ci              # Format, lint, test, docs
cargo make bench           # Rust pipeline benchmarks
cargo make check           # Type-check all crates (including excluded)
cargo make clippy          # Lint all crates (including excluded)

cargo make setup           # Install system deps + build C++ library
cargo make cpp-bench       # Run Rust vs C++ comparison benchmarks
cargo make cpp-test        # Run comparison tests
cargo make cpp-validate    # Run 2400+ C++ test suite with Rust backend

See Makefile.toml for the full list of tasks.

The minimum supported Rust version is 1.91.

• tonarino/webrtc-audio-processing -- Rust bindings to the C++ WebRTC AudioProcessing module. Sonora's examples are based on theirs. If you need the battle-tested C++ implementation with a Rust wrapper, use tonarino; if you want a pure-Rust solution with no C++ dependency, use sonora.
• pulseaudio-wap -- The original extracted C++ code

1. Google's libwebrtc -- The original AudioProcessing module is developed as part of the WebRTC Native Code project at Google.

2. PulseAudio's webrtc-audio-processing -- Arun Raghavan and contributors extracted the AudioProcessing module into a standalone library with a Meson build system, making it usable outside of the full WebRTC stack. This packaging is used by PulseAudio, PipeWire, and other Linux audio projects.

3. M145 upgrade and test expansion -- The C++ codebase was updated to WebRTC M145 (branch-heads/7632), the full upstream test suite (2400+ tests) was ported, and the build was upgraded to C++20. (AI assisted)

4. Sonora:Rust port -- The C++ implementation was ported to pure Rust, producing this set of crates. The port includes the full SIMD (SSE2, AVX2, NEON) optimizations and the FFTs, as well as a C API for FFI integration. The full C++ test suite passes against this Rust version. (AI assisted)

All crates in this repository are licensed under BSD-3-Clause.