rlx-aec

Pure Rust acoustic echo cancellation at 16 kHz mono — FDAF overlap-save NLMS (rustfft via rlx-fft) plus optional per-bin RLX residual suppression.

Pre-ASR front-end: feed cleaned PCM to Whisper / Voxtral unchanged.

CLI

cargo run -p rlx-aec --release -- \
  --mic-wav echoed_mic.wav \
  --ref-wav speaker_ref.wav \
  --out-wav cleaned.wav

Flags: --step-size, --max-delay-ms, --no-residual.

Library

use rlx_aec::{AecConfig, AecSession};

let mut aec = AecSession::new(AecConfig::default())?;
let clean = aec.process(&mic_chunk, &far_end_chunk)?;

Duplex: aec.push_reference(tts_pcm) on playback; aec.process_mic(mic_chunk) on capture.

Bench

just test-aec
just bench-aec
just bench-aec-parity   # Rust echo_bench + Python NLMS baseline

Weights

Embedded weights/residual_aec.safetensors (identity mask v1). Regenerate:

python3 scripts/aec_synthetic_dataset.py --out-dir /tmp/aec_dataset
python3 scripts/aec_train_residual_export.py --dataset-dir /tmp/aec_dataset \
  --out crates/rlx-aec/weights/residual_aec.safetensors

Tests