We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

eess

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Audio and Speech Processing

Title: Relaxed Attention: A Simple Method to Boost Performance of End-to-End Automatic Speech Recognition

Abstract: Recently, attention-based encoder-decoder (AED) models have shown high performance for end-to-end automatic speech recognition (ASR) across several tasks. Addressing overconfidence in such models, in this paper we introduce the concept of relaxed attention, which is a simple gradual injection of a uniform distribution to the encoder-decoder attention weights during training that is easily implemented with two lines of code. We investigate the effect of relaxed attention across different AED model architectures and two prominent ASR tasks, Wall Street Journal (WSJ) and Librispeech. We found that transformers trained with relaxed attention outperform the standard baseline models consistently during decoding with external language models. On WSJ, we set a new benchmark for transformer-based end-to-end speech recognition with a word error rate of 3.65%, outperforming state of the art (4.20%) by 13.1% relative, while introducing only a single hyperparameter.
Comments: Accepted at ASRU 2021, code contributed to this http URL
Subjects: Audio and Speech Processing (eess.AS); Computation and Language (cs.CL); Machine Learning (cs.LG); Sound (cs.SD)
Cite as: arXiv:2107.01275 [eess.AS]
  (or arXiv:2107.01275v2 [eess.AS] for this version)

Submission history

From: Timo Lohrenz [view email]
[v1] Fri, 2 Jul 2021 21:01:17 GMT (143kb)
[v2] Wed, 15 Dec 2021 11:10:05 GMT (299kb)

Link back to: arXiv, form interface, contact.