Source-linked AI summary
A Lightweight Hybrid Framework for Acoustic Echo Cancellation and Suppression via Constrained Recursive Training
Huawei Zhang, Rilin Chen, Hao Zhang, Meng Yu, Dong Yu
TL;DR
AEC seeks to remove echoes that can disrupt speech communication, while hybrid methods still face trade-offs in adaptability, stability, and efficiency. The paper combines a learnable NLMS filter with a jointly trained network and constrained recursive training. Its frameworks outperform baselines, with the proposed recursive framework reporting the highest ERLE and PESQ among hybrid methods and using 0.253 M parameters.
Problem
Existing hybrid methods face challenges balancing adaptability, stability, and efficiency, alongside gaps in filter design, joint training, training alignment, and model size.
Source: §1 · p. 1Method
Results
The proposed frameworks outperform baselines; the recursive framework achieves the highest ERLE and PESQ among hybrid methods with 0.253 M parameters.
Source: §6 · p. 4, §5.3 · p. 4Takeaways & Limitations
The framework combines adaptive filtering and neural post-processing in a lightweight implementation, with reported improvements across signal-quality and intelligibility measures.
Source: §1 · p. 1, §5.3 · p. 4, §6 · p. 4
Abstract
from arXiv · showhide
We propose a hybrid framework that combines deep learning with adaptive filtering for acoustic echo cancellation (AEC) and suppression. It integrates a multi-output neural network to jointly update filter parameters and perform post-processing. A learnable variant of the normalized least mean squares (NLMS) filter is introduced, incorporating a variable step size and a variable transition factor that are recursively updated for dynamic adaptation. In addition, an enhanced near-end speech estimate is generated from the filter output to suppress the residual echoes. To ensure training stability, we propose a constrained recursive training strategy that aligns with the adaptive nature of the filter, where an upper-bound constraint is introduced on the filter output. For efficiency, the framework adopts a lightweight implementation with only 253 k parameters. Experimental results show that the framework outperforms the baselines in terms of both signal quality and speech intelligibility.
1. INTRODUCTION
AEC removes echoes caused by acoustic coupling, but existing filters and learned approaches each have practical limitations. The paper addresses gaps in hybrid methods with a learnable NLMS filter, joint neural updates and post-processing, constrained recursive training, and a lightweight implementation.
1. INTRODUCTION
- AEC removes echoes caused by acoustic coupling between loudspeakers and microphones, which can disrupt conversations.The paper frames AEC as a fundamental component of modern speech communication systems.Source: §1 · p. 1
- Existing hybrid methods still struggle to balance adaptability, stability, and efficiency.The paper identifies further opportunities to improve these properties.Source: §1 · p. 1
- The paper identifies four gaps: filter design, separately trained networks, offline-training mismatch, and large network size.Offline training can mismatch the filter’s adaptive nature; complex nonlinear modeling contributes to parameter count.Source: §1 · p. 1
- The proposed VS-TF NLMS filter uses variable step size and transition factor for dynamic adaptation.It is designed to integrate strengths of NLMS and Kalman filters.Source: §1 · p. 1
- A single neural network jointly updates filter parameters and performs post-processing.This replaces the approach of combining two separately trained networks identified as a gap.Source: §1 · p. 1
- Constrained recursive training bounds the filter output and feeds it recursively into the network.The strategy is proposed to mitigate mismatch between offline training and online adaptation.Source: §1 · p. 1
2. ACOUSTIC ECHO CANCELLATION AND SUPPRESSION
The section formulates AEC as extracting near-end speech from a microphone signal containing echo, speech, and noise. It reviews adaptive-filter and learned approaches, including their performance, computational, and training challenges.
2.1. Problem Formulation
- The microphone signal is modeled as the sum of acoustic echo, near-end speech, and background noise.Using the far-end signal as reference, the goal is to extract the near-end speech from the received microphone signal.Source: §2.1 · p. 1
2.2. Related work
- NLMS is computationally efficient and easy to implement, but its performance is generally limited in severe scenarios.The error signal estimates near-end speech, and the filter update uses the reference-signal buffer and step size.Source: §2.2 · p. 2
- Kalman-filter performance depends on accurate covariance estimates, while deep-learning covariance estimation adds computational cost.A learnable Kalman variant reported comparable AEC performance at lower calculation cost than learned covariance estimation, but lacked theoretical support and interpretability.Source: §2.2 · p. 2
- Recursive training for acoustic howling suppression may be unstable; an AEC strategy estimates filter parameters but has difficulty handling nonlinear echo.The cited AEC strategy focuses on filter-parameter estimation.Source: §2.2 · p. 2
3. THE PROPOSED HYBRID FRAMEWORK
A single neural network jointly updates the adaptive filter and performs post-processing for AEC and suppression. The VS-TF NLMS filter provides dynamic adaptation, and post-processing suppresses residual echo in its output.
3. THE PROPOSED HYBRID FRAMEWORK
- One neural network uses microphone, reference, and filter-error features to produce adaptive-filter step size and transition factor, plus a post-processing ratio mask.Its inputs also include the previous RNN state, and it outputs the current state.Source: §3 · p. 2
3.1. Variable step-size and transition-factor (VS-TF) NLMS
- The VS-TF NLMS filter uses neural estimates of variable step size and transition factor to update the filter; estimating these parameters conveniently remains a challenge.The filter error signal is an initial estimate of near-end speech.Source: §3.1 · p. 2
3.2. Post-processing
- Neural post-processing estimates a complex ratio mask to suppress residual echo in the filter error signal and produce the final near-end speech estimate.The framework uses the filter error as the input to this post-processing stage.Source: §3.2 · p. 2
4. TRAINING STRATEGY
Training recursively applies the network's previous outputs to update the adaptive filter and feeds its current output back into the network. An upper-bound constraint limits the filter output during training and streaming inference, alongside sub-losses for the filter and post-processing outputs.
4.1. Recursive training
- Recursive training uses previous network outputs to update the current filter, then feeds its current error output back into the network.The error signal is included to provide information about the adaptive process, which offline-trained models described in the passage did not consider.Source: §4.1 · p. 3
4.2. Upper-bound constraint
- The filter output is constrained not to exceed the microphone signal's amplitude, limiting neural-network input values to avoid training instability.The same constraint is applied during streaming inference.Source: §4.2 · p. 3
4.3. Loss function
- The filter and post-processing outputs have separate sub-losses; the stated loss terms are SI-SNR, Echo-Aware Loss, and MAE, weighted by α1 = 0.04 and α2 = 0.005.The overall loss is used for backpropagation during training.Source: §4.3 · p. 3
5. EXPERIMENTAL STUDY
The study evaluates the proposed framework on simulated training and validation data and an AEC Challenge evaluation set. Results report strong performance from the recursive framework, including hybrid-baseline comparisons and spectrogram analysis.
5.1. Data preparation
- Evaluation uses 100 AEC Challenge utterances with SER from −10 dB to 10 dB; simulated 8-second mixtures draw speech, noise, and room responses from named datasets.The study reports SISNR, SDR, ERLE, and PESQ; the evaluation sampling frequency is 16000 Hz.Source: §5.1 · p. 3, §5.2 · p. 4
5.2. Experimental setting
- Comparisons include classical NLMS and Kalman filters, NLMSNet and KalmanNet, VS-TF NLMS, half-recursive, and recursive frameworks; neural-network training uses the same loss for proposed methods and baselines.VS-TF NLMS estimates its step size and transition factor with a single-layer RNN; half-recursive uses offline GTCRN post-filtering, while recursive uses GTCRN in the proposed framework.Source: §5.2 · pp. 3–4
5.3. Experimental results
- On validation data, the recursive framework adds about 11.5 dB SISNR and roughly 10 dB SDR over VS-TF NLMS, with nearly 0.4 higher PESQ; it achieves the best overall performance with a compact model.The VS-TF NLMS filter improves SI-SNR by approximately 1.2 dB and SDR by about 0.9 dB over classical NLMS; its parameter count is 0.067 M, compared with 0.168 M for half-recursive and 0.253 M for recursive.Source: §5.3 · p. 4
- The recursive framework achieves the best performance among hybrid methods on evaluation-set ERLE and PESQ, using 0.253 M parameters versus 0.888 M for NLMSNet and 0.977 M for KalmanNet.ERLE evaluates single-talk scenarios and PESQ evaluates double-talk scenarios.Source: §5.3 · p. 4
- The spectrogram comparison shows stronger echo removal by the recursive framework than by NLMSNet and KalmanNet, while KalmanNet and the recursive framework better preserve speech above 2 kHz than NLMSNet.Fig. 2 presents the microphone signal, near-end speech, and outputs from NLMSNet, KalmanNet, and the proposed recursive framework.Source: §5.3 · p. 4, Figure 2 · p. 4
6. CONCLUSION
The conclusion presents a lightweight hybrid AEC and suppression framework built around a VS-TF NLMS filter and constrained recursive training.
6. CONCLUSION
- The proposed frameworks with the VS-TF NLMS filter outperform the baselines; the recursive framework uses 253 k parameters, while the half-recursive variant uses 168 k.The half-recursive variant trains part of the neural network recursively and the rest offline.Source: §6 · p. 4