How to Objectively Quantify Acoustic Echo Cancellation Performance

Introduction

Acoustic Echo Cancellation (AEC) is a common audio signal processing technique used to suppress echo in voice communication. Traditional subjective evaluations are influenced by human bias, so an objective, data‑based method is needed to quantify AEC performance.

Background

With the rise of voice communication, echo and noise degrade call quality. Subjective methods suffer from high subjectivity, poor repeatability, and low efficiency, especially across many platforms and devices. An objective evaluation is crucial for improving AEC algorithms.

Test Environment

Hardware List

Test Network

The test setup places device A in front of an artificial head. Device B sends speech to A while the head plays a near‑end signal and a speaker simulates environmental noise. The final signal captured by device B is analyzed.

Core Metrics

Subjective concerns are translated into objective indicators:

Initial leak echo (e.g., echo heard when joining a room)

Residual echo magnitude

Residual echo stability

Suppression of near‑end speech or background noise on the far end (single‑talk)

Leak echo during simultaneous speaking (double‑talk)

These are quantified using open‑source and proprietary methods.

Test Flow and Scenario Classification

Single‑Talk Test

Device B sends a signal to A; A plays it back, captures the echo, processes it with AEC, sends it back to B, and B records the result for metric calculation. Evaluation can be reference‑free (analyzing echo magnitude and stability) or reference‑based (using a near‑end reference microphone).

Double‑Talk Test

Two rounds are performed: first only the far‑end signal, then both far‑end and near‑end speech. Metrics such as double‑talk cut, residual echo, intelligibility, and MOS are combined to assess performance.

Model Scoring

For each device model, three scores are derived:

Raw values : absolute results from core metrics.

Mapped scores : non‑linear conversion to a 0‑100 scale, considering subjective experience and penalties.

Final score : weighted aggregation of mapped scores, with weights adjusted per usage scenario.

Result Analysis

Comparing two SDK versions on a specific device shows that version B outperforms version A in residual echo level, stability, and double‑talk intelligibility. Objective data align with subjective listening tests, confirming the validity of the evaluation method.

Environment Showcase

Using the cloud‑based API, more than ten devices can be tested simultaneously with a single click, automatically generating comprehensive reports.

Outcome and Outlook

Continuous algorithm improvements, driven by objective evaluation, have enhanced the clarity and naturalness of audio calls. The quantitative AEC assessment method merges scientific rigor with practical relevance, and future work will focus on further innovation and optimization to deliver superior audio services.