Abstract
Echo path delay (or ref-delay) estimation is a big challenge in acoustic echo cancellation. Different devices may introduce various ref-delay in practice. Ref-delay inconsistency slows down the convergence of adaptive filters, and also degrades the performance of deep learning models due to 'unseen' ref-delays in the training set. In this paper, a multi-task network is proposed to address both ref-delay estimation and echo cancellation tasks. The proposed architecture consists of two convolutional recurrent networks (CRNNs) to estimate the echo and enhanced signals separately, as well as a fully-connected (FC) network to estimate the echo path delay. Echo signal is first predicted, and then is combined with reference signal together for delay estimation. At the end, delay compensated reference and microphone signals are used to predict the enhanced target signal. Experimental results suggest that the proposed method makes reliable delay estimation and outperforms the existing state-of-the-art solutions in inconsistent echo path delay scenarios, in terms of echo return loss enhancement (ERLE) and perceptual evaluation of speech quality (PESQ). Furthermore, a data augmentation method is studied to evaluate the model performance on different portion of synthetical data with artificially introduced ref-delay.
Introduction
Acoustic Echo Cancellation (AEC) is the process of removing the "echo" created when a microphone picks up sound from a speaker and sends it back to the far-end caller. A major hurdle in AEC is echo path delay (ref-delay). Because different hardware (phones, PCs, tablets) and software stacks introduce varying amounts of latency, the "reference" signal (what the speaker plays) and the "microphone" signal (what the mic hears) are often out of sync. Traditional adaptive filters struggle to converge when delays are inconsistent, and standard deep learning models often fail when they encounter a delay they didn't see during training. This paper proposes a single, unified network that can estimate the delay on the fly and use that information to clean the audio more effectively.
Proposed Multi-task Architecture
The authors propose a multi-task network composed of three distinct sub-modules designed to decouple delay estimation from signal enhancement:
- Echo Estimation CRNN: A Convolutional Recurrent Network (CRNN) that takes the microphone waveform and predicts the echo signal. By predicting a "clean" echo first, the model avoids performing cross-correlation on the raw microphone signal, which is often contaminated by near-end speech.
- Delay Estimation FC Network: This fully-connected network takes the estimated echo and the original reference signal. It predicts the delay as a "one-hot" vector, categorizing the delay into 10-sample buckets up to 400 samples.
- Enhancement CRNN: Once the delay is estimated, the reference signal is shifted to align with the microphone signal. This "delay-compensated" reference and the microphone signal are fed into a second CRNN to produce the final, clean near-end speech.
The model uses Focal Loss for the delay estimation task to handle data imbalance and Mean Square Error (MSE) for the speech enhancement task. By compensating for the delay before the final enhancement stage, the second CRNN can focus entirely on modeling the Room Impulse Response (RIR) and suppressing the echo rather than wasting capacity on temporal shifts.
Experimental Results
The proposed multi-task model was evaluated using the TIMIT dataset for speech and real-world Room Impulse Responses (RIRs) from the RWCP database. Synthetic augmentation was used to introduce delays ranging from 0 to 400 samples.
Performance in Mismatched Scenarios
When trained on zero-delay data but tested on delayed data, standard CRNN performance dropped significantly (PESQ from 2.70 to 2.35). In contrast, the multi-task model maintained a PESQ of 2.75 and an ERLE (Echo Return Loss Enhancement) of 50.9 dB, demonstrating robust performance even in "unseen" delay scenarios.
Delay Estimation Accuracy
The model achieved high precision in delay estimation. As shown in the results, over 80% of estimates fell within 10 samples of the true delay in simple scenarios, and over 50% in complex RIR scenarios.
Conclusion and Future Work
The multi-task approach successfully addresses the challenge of inconsistent echo path delays. However, some limitations remain. The current model assumes a linear echo path, whereas real-world speakers often introduce non-linear distortions. Additionally, the delay estimation resolution is limited to 10-sample increments due to the maxpooling layers used for efficiency. Future work may explore higher precision estimation and modeling non-linear echo paths for improved performance in diverse hardware environments.
Summary updated