OSLEC Simulation Environment

Overview

A user-space harness for exercising the OSLEC echo canceller without touching the original kernel-oriented sources. The workflow builds echo.c as a shared library, exposes it through ctypes, and ships helper modules for signal generation, echo-path modelling, and performance analysis.

Prerequisites

• GCC toolchain for building the shared library
• Python 3.9+ with venv
• numpy, scipy, and matplotlib (install via the provided requirements)

Building liboslec.so

cd simulation
make     # produces liboslec.so next to the Python sources

The build relies on lightweight stub headers in simulation/include/linux to satisfy the kernel dependencies used by echo.c.

Python Environment Setup

cd simulation
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Running a Simulation

cd simulation
python3 test_suite.py --duration 6 --sample-rate 8000 --noise-snr-db 35

Optional flags:

• --double-talk enables simultaneous near-end speech
• --plot opens matplotlib visualisations
• --taps, --delay-ms, --attenuation-db tune echo-canceller parameters
• --verbose prints debug logging to help diagnose convergence

A typical run prints ERLE statistics and convergence timing. Use --help for the full CLI.

Choosing Tap Length

For good convergence make sure the tap count spans the entire simulated echo path. With an echo delay of delay_ms and a tail of roughly tail_taps additional samples, use:

taps >= sample_rate * delay_ms / 1000 + tail_taps

Example: --delay-ms 32 with a 128-tap tail at 8 kHz needs ~384 taps.

Module Layout

• oslec_wrapper.py – ctypes wrapper with helpers to convert PCM frames
• signal_generator.py – synthetic speech, tones, and noise builders
• echo_simulator.py – impulse-response modelling and noise injection
• scenarios.py – declarative scenario definitions (stationary, time-variant, double-talk, tone interference)
• analyzer.py – ERLE, convergence, and power metrics
• benchmark.py – batch comparison tooling with hyperparameter sweeps
• test_suite.py – configurable scenarios + optional plotting frontend

Benchmark Results (3 s clip, 384 taps)

We evaluated OSLEC against lighter-weight NLMS, IPNLMS, and MDF references across four scenarios. Figures are saved under simulation/plots/.

Key takeaways:

• OSLEC still delivers the highest steady-state ERLE, particularly for stationary and time-varying echo paths, thanks to its dual-path update and auxiliary NLP/CNG logic.
• IPNLMS converges significantly faster and yields the lowest residual noise in stationary conditions, but its mean ERLE trails OSLEC. It is a good candidate when quick adaptation is more important than absolute suppression.
• NLMS benefits from tuned step sizes (μ ≈ 0.8) yet remains a baseline option for minimal complexity. MDF in its simplified form requires further optimisation before it can compete.
• Tone interference remains difficult for the lightweight algorithms; the extra processing inside OSLEC keeps it marginally ahead.

Regenerate the full benchmark (plots plus summary tables) with:

cd simulation
source .venv/bin/activate
BENCHMARK_PLOT_DIR=plots MPLBACKEND=Agg BENCHMARK_SAVE=1 \
  python3 benchmark.py --duration 3 --taps 384

Next Steps

• Extend the scenario suite with fax-tone stress tests and adaptive-tail sweeps
• Integrate PESQ or POLQA tooling for perceptual quality scoring
• Hook benchmark automation into CI for regression tracking