SAGRNN: Self-Attentive Gated RNNs for Binaural Speaker Separation with Interaural Cue Preservation

Ke Tan1, Buye Xu2, Anurag Kumar2, Eliya Nachmani3, 4, Yossi Adi3

1The Ohio State University, 2Facebook Reality Labs, 3Facebook AI Research, 4Tel-Aviv University


In this post, we present a novel framework called multiple-input multiple-output self-attentive gated recurrent neural network (MIMO SAGRNN) for binaural speaker separation, which produces a binaural estimate for each of the target speakers with preserved interaural cues. This approach was proposed in our paper titled "SAGRNN: Self-Attentive Gated RNNs for Binaural Speaker Separation with Interaural Cue Preservation", accepted by IEEE Signal Processing Letters.

In this study, the goal of binaural speaker separation is to separate \(C\) concurrent binaural speech signals \(\mathbf{s}_1, \dots, \mathbf{s}_C\) from a sound mixture \(\mathbf{y}\), where \(\mathbf{y} = \sum_{c=1}^C \mathbf{s}_c\) (noise-free condition) or \(\mathbf{y} = \sum_{c=1}^C \mathbf{s}_c + \sum_{c_n=1}^{C_n} \mathbf{n}_{c_n} \) (noisy condition). Here \(\mathbf{n}_{c_n}\) denotes the \(c_n\)-th binaural noise signal and \(C_n\) the number of noise sources. We assume that there is no room reverberation in the acoustic environment.

The SAGRNN network architecture extends the gated RNN in [1] by additionally incorporating self-attention mechanisms and dense connectivity. MIMO SAGRNN is then derived from a single-input single-output (SISO) SAGRNN by first extending the SISO SAGRNN into a multiple-input single-output (MISO) layout, in which we create two encoders, one for the reference ear input and the other for the non-reference ear input. This MISO SAGRNN estimates the separated signals in the reference ear. The MIMO system is formulated by alternately treating each ear as the reference ear, yielding estimates for both ears in a symmetric manner.

Our experimental results show that the proposed approach significantly outperforms a recent binaural separation approach (i.e. MIMO TasNet [2]) in terms of ΔSDR, ΔSNR, ESTOI and PESQ. Moreover, our approach effectively preserves the auditory spatial cues of talkers.

For full details, see our paper.

Fig. 1. Diagram of SISO SAGRNN. The dotted lines represent the procedures that exist only during training.

Fig. 2. Diagram of the SA-MULCAT block. The dotted lines indicate that the procedures exist in all SA-MULCAT blocks except the first one. The symbol \(\bigotimes\) represents the element-wise multiplication, and \(\bigoplus\) the element-wise addition.

Fig. 3. (a) Diagram of the self-attention block. (b) Overview of the MIMO separation system with an underlying MISO system.



Here are some (binaural) audio samples for listening:

I. Male-Female Mixtures

Sample IDs Unprocessed Mixture MIMO TasNet MIMO SAGRNN

II. Male-Male Mixtures

Sample IDs Unprocessed Mixture MIMO TasNet MIMO SAGRNN

III. Female-Female Mixtures

Sample IDs Unprocessed Mixture MIMO TasNet MIMO SAGRNN

IV. A "Crossfading" Demo

The following demo crossfades between two binaural speech signals separated by our MIMO SAGRNN. You can first click the "Play/Stop" button to start repeatly playing the audios, and then drag the scroll box to adjust the weighting factors for crossfading (or remixing) the two speakers. You need to wear a headphone for binaural hearing, and we expect you to perceive the different directions of the two speech sources. This demo is created to help demonstrate that the interaural cues are preserved in the separated speech signals by our approach.
Speaker 1
Speaker 2

As you may hear, the female speaker is in the left half-space, and the male speaker is in the right half-space. We create another demo for comparison, in which each speaker only has the signal for the left-ear channel (i.e. mono audio).
Speaker 1
Speaker 2

Now we are unable to perceive the directions of the speech sources due to the absence of interaural cues.


[1] E. Nachmani, Y. Adi, and L. Wolf. Voice separation with an unknown number of multiple speakers. In International Conference on Machine Learning, 2020.

[2] C. Han, Y. Luo, and N. Mesgarani. Real-time binaural speech separation with preserved spatial cues. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 6404–6408. IEEE, 2020.


I. The CIPIC HRTF Database


Copyright Notice: THE REGENTS OF THE UNIVERSITY OF CALIFORNIA MAKE NO REPRESENTATION OR WARRANTIES WITH RESPECT TO THE CONTENTS HEREOF AND SPECIFICALLY DISCLAIM ANY IMPLIED WARRANTIES OR MERCHANTAB ILITY OR FITNESS FOR ANY PARTICULAR PURPOSE. Further, the Regents of the University of California reserve the right to revise this software and/or documentation and to make changes from time to time in the content hereof without obligation of the Regents of the University of California to notify any person of such revision or change.

Citation: V. R. Algazi, R. O. Duda, D. M. Thompson, and C. Avendano. The cipic hrtf database. In Proceedings of the 2001 IEEE Workshop on the Applications of Signal Processing to Audio and Acoustics (Cat. No. 01TH8575), pages 99–102. IEEE, 2001.



License(s): LDC User Agreement for Non-Members

Citation: J. Garofolo, D. Graff, D. Paul, and D. Pallett. Csr-i (wsj0) complete ldc93s6a. Web Download. Philadelphia: Linguistic Data Consortium, 83, 1993.

III. Freesound Noises


Attribution Noncommercial: