Abstract
The user-centric residual-echo suppression (URES) framework accepts a user-operating point (UOP) comprising two metrics: the residual-echo suppression level (RESL) and the desired-speech maintained level (DSML). It produces several RES-system predictions with different UOP estimates, and the prediction with the highest acoustic-echo cancellation mean-opinion score (AECMOS) within the UOP tolerance becomes the output. Despite showing promising results, its high computational burden limits applicability. This paper introduces an efficient URES (E-URES) framework, which reduces computational costs in the final stage of the URES pipeline by minimizing the number of AECMOS computations. A lightweight neural network learns the relation between the UOP estimates and their corresponding AECMOS values by feeding the network various acoustic signals. During inference, the framework uses the three highest AECMOS predictions within the tolerance limit of the UOP to determine which outcomes to carry the actual AECMOS computations. Using 60 hours of data, average results show that the E-URES reduces 90% of the computational cost with negligible performance reduction.
Original language | English |
---|---|
Title of host publication | 2024 18th International Workshop on Acoustic Signal Enhancement, IWAENC 2024 - Proceedings |
Publisher | Institute of Electrical and Electronics Engineers Inc. |
Pages | 364-368 |
Number of pages | 5 |
ISBN (Electronic) | 9798350361858 |
DOIs | |
State | Published - 2024 |
Externally published | Yes |
Event | 18th International Workshop on Acoustic Signal Enhancement, IWAENC 2024 - Aalborg, Denmark Duration: 9 Sep 2024 → 12 Sep 2024 |
Publication series
Name | 2024 18th International Workshop on Acoustic Signal Enhancement, IWAENC 2024 - Proceedings |
---|
Conference
Conference | 18th International Workshop on Acoustic Signal Enhancement, IWAENC 2024 |
---|---|
Country/Territory | Denmark |
City | Aalborg |
Period | 9/09/24 → 12/09/24 |
Bibliographical note
Publisher Copyright:© 2024 IEEE.
Keywords
- AECMOS
- Residual-echo suppression
- deep learning
- low compute
- user-centric