REVIEW 1 cited by
Ultra-Low Latency Speech Enhancement - A Comprehensive Study
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Ultra-Low Latency Speech Enhancement - A Comprehensive Study
read the original abstract
Speech enhancement models should meet very low latency requirements typically smaller than 5 ms for hearing assistive devices. While various low-latency techniques have been proposed, comparing these methods in a controlled setup using DNNs remains blank. Previous papers have variations in task, training data, scripts, and evaluation settings, which make fair comparison impossible. Moreover, all methods are tested on small, simulated datasets, making it difficult to fairly assess their performance in real-world conditions, which could impact the reliability of scientific findings. To address these issues, we comprehensively investigate various low-latency techniques using consistent training on large-scale data and evaluate with more relevant metrics on real-world data. Specifically, we explore the effectiveness of asymmetric windows, learnable windows, adaptive time domain filterbanks, and the future-frame prediction technique. Additionally, we examine whether increasing the model size can compensate for the reduced window size, as well as the novel Mamba architecture in low-latency environments.
Forward citations
Cited by 1 Pith paper
-
Low-latency Assistive Audio Enhancement for Neurodivergent People
Among DSP and ML audio enhancement approaches evaluated on trigger-sound mixtures, Dynamic Range Compression (DRC) attenuates distressing sounds most effectively in both objective metrics and a neurodivergent listening test.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.