Python Acoustic Echo Cancellation Library

Overview

This repository implements classic adaptive filters—LMS, NLMS, and RLS—geared toward real-time Acoustic Echo Cancellation (AEC), noise reduction, and general audio DSP.

Filter	Highlights
LMS	Lightweight; now supports a safe mode that clips extreme error values to prevent numeric overflows.
NLMS	Normalises step size per block ⇒ faster, stabler convergence.
RLS	Uses an inverse correlation matrix for very rapid adaptation.

---

Installation

pip install numpy scipy matplotlib sounddevice   # sounddevice is needed for the live demo

⚠️ The package is source-only for now—just clone and import:

git clone https://github.com/Keyvanhardani/Python-Acoustic-Echo-Cancellation-Library.git
cd Python-Acoustic-Echo-Cancellation-Library

---

Quick Start

1 · Run the built-in demo

python lms.py     # picks the best μ and prints the total error

2 · Live AEC over mic + speaker

python mic_demo.py    # press Ctrl-C to stop

You’ll hear your own voice while the loud-speaker echo is strongly attenuated.

---

API Examples

LMS (with safe=True)

from lms import lms_filter
import numpy as np

d = np.random.randn(10_000).astype(np.float32)  # desired signal
u = np.random.randn(10_000).astype(np.float32)  # reference (echo path)
f0 = np.zeros(1024, dtype=np.float32)           # initial coeffs
mus = [1e-4, 5e-4, 1e-3]

e, f_adapt, best_mu = lms_filter(
        desired_signal=d,
        reference_input=u,
        filter_coeff=f0,
        step_sizes=mus,
        safe=True        # clips |e| ≤ 1e4 → no RuntimeWarnings
)

NLMS

from nlms import nlms_filter
e, f_adapt = nlms_filter(d, u, f0, mu=0.5)

RLS

from rls import rls_filter
e, f_adapt = rls_filter(d, u, f0, reg_param=0.1, lambda_=0.98)

---

Batch AEC: WAV → filtered WAV

import scipy.io.wavfile as wav
from lms import lms_filter
import numpy as np

fs, y = wav.read("input.wav")        # y is int16
x = y.astype(np.float32) / 32768.0   # scale to ±1

e, *_ = lms_filter(x, x, np.zeros(1024), [5e-4], safe=True)

wav.write("output.wav", fs, np.int16(e / np.max(np.abs(e)) * 32767))

---

Visualising the Result

import matplotlib.pyplot as plt
plt.plot(x,   label="Original")
plt.plot(e,   label="AEC output", alpha=0.7)
plt.legend(); plt.show()

---

Live Demo Options

Flag	Description	Default
BLOCK	FFT/block length (also filter length)	1024
FS	Sample rate	48 000 Hz
MU	Step size	5 × 10⁻⁴

Adjust them directly at the top of mic_demo.py.

---

Contributing

1. Fork & create a feature branch.
2. Run python lms.py and python mic_demo.py—there must be zero warnings.
3. Submit a PR with a clear description.

---

License

MIT