E-Gaborformer: Strictly Causal Speech Recognizer with Constant Streaming Memory
arXiv cs.LG (Machine Learning) RSS Feed ·
What to know
A new arXiv paper presents E-Gaborformer, a strictly causal encoder combining Gated DeltaNet-2 recurrent layers with sliding-window attention and fixed causal Gabor atoms. It maintains a constant streaming state of about 19 MB per call at production scale, versus over 1 GB for a same-size causal-attention encoder after ten minutes. In production, a 0.6B-parameter Hebrew and English streaming transducer built on it achieves 28–31% lower WER than NVIDIA's Nemotron 3.5 ASR Streaming on English voice-agent traffic, and up to 20% lower Hebrew WER than the previous system, with 2.6x faster training via a two-class RNN-T lattice formulation.
Source and evidence
arXiv cs.LG (Machine Learning) RSS FeedOriginal language · und ·Original author: Nir Krakowski, Tamir Cohen, Ohad Zohar, Chen Zarfati · arXiv Readzer provides an abridged translation. CC0.Source licence
Why it matters by industry
- Technology
Relevant to software and digital infrastructure: introduces a new speech recognition architecture with constant memory and lower WER, directly impacting voice-agent systems and streaming ASR deployment.
Reviewed summary. The complete article remains at its original source.