The REVERB Challenge: A Benchmark Task for Reverberation-Robust ASR Techniques

Keisuke Kinoshita, Marc Delcroix, Sharon Gannot, Emanuël A. P. Habets, Reinhold Haeb-Umbach, Walter Kellermann, Volker Leutnant, Roland Maas, Tomohiro Nakatani, Bhiksha Raj, Armin Sehr, Takuya YoshiokaNew Era for Robust Speech Recognition, Exploiting Deep Learning 2017 · OpenReview

Abstract

The REVERB challenge is a benchmark task designed to evaluate reverberation-robust automatic speech recognition techniques under various conditions. A particular novelty of the REVERB challenge database is that it comprises both real reverberant speech recordings and simulated reverberant speech, both of which include tasks to evaluate techniques for 1-, 2-, and 8-microphone situations. In this chapter, we describe the problem of reverberation and characteristics of the REVERB challenge data, and finally briefly introduce some results and findings useful for reverberant speech processing in the current deep-neural-network era.
BibTeX
@inproceedings{kinoshita2017reverb,
  author = {Keisuke Kinoshita and Marc Delcroix and Sharon Gannot and Emanuël A. P. Habets and Reinhold Haeb-Umbach and Walter Kellermann and Volker Leutnant and Roland Maas and Tomohiro Nakatani and Bhiksha Raj and Armin Sehr and Takuya Yoshioka},
  title = {{The REVERB Challenge: A Benchmark Task for Reverberation-Robust ASR Techniques}},
  booktitle = {New Era for Robust Speech Recognition, Exploiting Deep Learning},
  year = {2017},
  url = {https://openreview.net/forum?id=D4Rl8efLmsH},
}

Loading related content...

Figure