Deep Model with Built-in Cross-Attention Alignment for Acoustic Echo Cancellation

Evgenii Indenbom

Ristea Nicolae Catalin

Ando Saabas

Tanel Parnamaa

Jegor Gužvin

Ross Cutler

Microsoft Corporation

Abstract

With recent research advances, deep learning models have become an attractive choice for acoustic echo cancellation (AEC) in real-time teleconferencing applications. Since acoustic echo is one of the major sources of poor audio quality, a wide variety of deep models have been proposed. However, an important but often omitted requirement for good echo cancellation quality is the synchronization of the microphone and far end signals. Typically implemented using classical algorithms based on cross-correlation, the alignment module is a separate functional block with known design limitations. In our work, we propose a deep learning architecture with built-in cross-attention-based alignment, which is able to handle unaligned inputs, improving echo cancellation performance while simplifying the communication pipeline. Moreover, we show that our approach achieves significant improvements for difficult delay estimation cases on real recordings from the AEC Challenge data set.

Official DEMO

Extended results

Method	FEST	DT Echo	DT Other	Inference time	Measurement
	MOS	MOS	MOS	per frame (ms)	Device
#1 Place AEC Challenge	4.34	4.36	4.23	2.62	i5-4300 CPU @1.9GHz
#2 Place AEC Challenge	4.44	4.44	3.90	4.385	Xeon(R) CPU E5-2640@2.50GHz
#1 Rank AEC Challenge	4.59	4.69	4.18	N/A	N/A
CRUSE (online)	4.55	4.42	4.07	0.216	i7 11370H@3.3 GHz
Align-CRUSE	4.67	4.45	4.07	0.218	i7 11370H@3.3 GHz

Reverb and Noise Suppression results

Method	SRR (dB)	SIG	BAK	OVRL	Inference time	Measurement device
Noisy	15.24	3.13	3.26	2.67	-	-
#1 Place AEC Challenge	35.87	4.39	4.59	4.11	2.62	i5-4300 CPU @1.9GHz
#2 Place AEC Challenge	34.79	3.78	4.21	3.52	4.385	Xeon(R) CPU E5-2640@2.50GHz
#1 Rank AEC Challenge	34.37	4.37	4.58	4.07	N/A	N/A
CRUSE (online)	36.54	4.41	4.54	4.11	0.216	i7 11370H@3.3 GHz
Align-CRUSE	37.47	4.39	4.55	4.10	0.218	i7 11370H@3.3 GHz

To have a more comprehensive view of our model, we tested the reverb and noise suppression capability on the blind near end test set from the AEC Challenge. We additionally reported speech-to-reverberation ratio (SRR), the DNSMOS P.835 Signal (SIG), background (BAK) and overall (OVRL). The SRR value is estimated in decibels with an internal deep model. In the above table we observe that Align-CRUSE has competitive results with state-of-the-art methods for both noise suppression and AEC. Regarding the reverb removal capacity, we observe that our model obtains the best results. This highlights that our model is able to jointly enhance the audio quality and cancel echoes.

Considering the inference time and the quality results, we conclude that Align-CRUSE is state-of-the-art method for low-complexity AEC, which also attains very good results for the reverb and noise reduction tasks.

AEC Challenge test set (FEST-GEN and FEST-HD)

Microphone:
Far end:
CRUSE (online):	AECMOS: 1.35, ERLE: 4.92
Align-CRUSE:	AECMOS: 4.83, ERLE: 70.55
		Align feature map

Microphone:
Far end:
CRUSE (online):	AECMOS: 3.07, ERLE: 31.48
Align-CRUSE:	AECMOS: 4.18, ERLE: 43.97
		Align feature map

Microphone:
Far end:
CRUSE (online):	AECMOS: 4.63, ERLE: 55.76
Align-CRUSE:	AECMOS: 4.73, ERLE: 74.74
		Align feature map

Microphone:
Far end:
CRUSE (online):	AECMOS: 4.44, ERLE: 15.91
Align-CRUSE:	AECMOS: 4.47, ERLE: 17.36
		Align feature map

Microphone:
Far end:
CRUSE (online):	AECMOS: 4.40, ERLE: 44.03
Align-CRUSE:	AECMOS: 4.48, ERLE: 46.57
		Align feature map

Synthetic (0.3-0.5)s delays test set (LD-M)

This test set is synthetically created, therefore we included the label delay and the estimated delay from the align feature map.

Microphone:
Far end:
CRUSE (online):	AECMOS: 2.68, ERLE: 12.99
Align-CRUSE:	AECMOS: 4.70, ERLE: 48.35
		Align feature map True delay: 305ms, Estimated average delay: 315ms

Microphone:
Far end:
CRUSE (online):	AECMOS: 3.45, ERLE: 23.75
Align-CRUSE:	AECMOS: 4.72, ERLE: 56.39
		Align feature map True delay: 355ms, Estimated average delay: 361ms

Synthetic (0.5-1.0)s delays test set (LD-H)

Microphone:
Far end:
CRUSE (online):	AECMOS: 1.97, ERLE: 1.00
Align-CRUSE:	AECMOS: 4.80, ERLE: 62.66
		Align feature map True delay: 916ms, Estimated average delay: 905ms

This test set is synthetically created, therefore we included the label delay and the estimated delay from the align feature map.

Microphone:
Far end:
CRUSE (online):	AECMOS: 2.80, ERLE: 0.88
Align-CRUSE:	AECMOS: 4.63, ERLE: 28.02
		Align feature map True delay: 763ms, Estimated average delay: 739ms

AEC Challenge Double-Talk

Microphone:
Far end:
Align-CRUSE:
		Align feature map

Microphone:
Far end:
Align-CRUSE:
		Align feature map

Additional information

We included samples only for Align-CRUSE and online CRUSE, since CRUSE tested on unaligned data is considerably worse (see Table 1 from paper) and global aligned CRUSE is not feasible for real-time applications.
The align feature map represents the delay distributions (between 0 and 1 second) for 20ms frames. More precisely, the vertical axis represents the probability delay distribution for a certain frame index.
For a better align feature map visualization, we included only the middle 500 frames from a sample.
The estimated average delay is computed by averaging the highest probability delay for each frame in the sample. It was reported only for synthetically data sets since for those samples we have the ground-truth.