Fig. 5
From: Strategies for distant speech recognitionin reverberant environments
Influence of extended training data. The figure plots the WER as a function of the training data used for a SimData and b RealData, for the development set. The results are given for different SE front-ends, i.e., no SE front-end (‘Distant’), 1ch WPE (‘WPE (1ch)’) and the combination of 8ch WPE with MVDR and DOLPHIN (‘WPE (8ch) + MVDR + DOL’). The left bars (‘Baseline 20 dB’) show the results for the baseline multi-condition training data set. The other bars show the results when the amount of training data is gradually increased by adding distant speech at SNRs of 10 and 15 dB (‘+10 dB, 15 dB’), the clean training data (‘+ clean’), and the training data recorded with a desktop microphone (‘desktop’)