# On TasNet for Low-Latency Single-Speaker Speech Enhancement

**Authors:** Morten Kolbæk, Zheng-Hua Tan, Søren Holdt Jensen, Jesper Jensen  
**arXiv:** [2103.14882](https://arxiv.org/abs/2103.14882)  
**Date:** 2021-03-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2103.14882v1/

## Key points

- TasNet is a time-domain audio separation network used for single-speaker speech enhancement.
- TasNet improves state-of-the-art performance in speech enhancement by separating target and noise signal components effectively.
- TasNet excels at separating interfering speech signals from noise due to its ability to learn an efficient inner-domain representation.
- Potential issues with large frame hops can affect TasNet's performance due to aliasing problems.
- Experimental simulations using speech signals contaminated by additive noise evaluate TasNet's effectiveness, with metrics like STOI, PESQ, and Scale-Invariant SDR used for assessment.
- TasNet is tested as a 2-speaker speech separation system using the WSJ0 speech corpus for training data and mean STOI, PESQ, and SI-SDR metrics for evaluation.
- TasNet shows promise for low-latency single-speaker speech enhancement applications by effectively separating target speech from various noise sources like modulated noise and interfering speech signals.

## Summary

, , , , 

The paper "On TasNet for Low-Latency Single-Speaker Speech Enhancement" delves into the use of TasNet, a time-domain audio separation network, for single-speaker speech enhancement. The study demonstrates that TasNet, known for its success in speech separation tasks, also improves state-of-the-art performance in speech enhancement. Notably, it excels at separating target and noise signal components from modulated noise sources such as speech. This is attributed to TasNet's ability to learn an efficient inner-domain representation, particularly in separating interfering speech signals from noise. However, the study also highlights potential issues with large frame hops that can affect TasNet's performance due to aliasing problems. To evaluate TasNet's effectiveness as an enhancement system, experimental simulations are conducted using speech signals contaminated by additive noise. Performance metrics such as STOI, PESQ, and Scale-Invariant SDR are used to assess its performance. Additionally, TasNet is tested as a 2-speaker speech separation system to validate its implementation against existing literature results. The WSJ0 speech corpus is used for training data and mean STOI, PESQ, and SI-SDR metrics are used to evaluate performance. Overall, the findings suggest that TasNet shows promise for low-latency single-speaker speech enhancement applications by effectively separating target speech from various noise sources like modulated noise and interfering speech signals. Future research may focus on addressing limitations related to large frame hops to enhance the overall efficiency of TasNet in real-world scenarios.