Data Augmentation for Training of Noise Robust Acoustic Models.

In this paper we analyse ways to improve the acoustic models based on deep neural networks with the help of data augmentation. These models are used for speech recognition in a priori unknown possibly noisy acoustic environment (with the presence of office or home noise, street noise, babble, etc.) and may deal with both the headset and distant microphone recordings. We compare acoustic models trained on speech corpora with artificially added noises of different origins and reverberation. At various test sets, word recognition accuracy improvement over the baseline model trained on clean headset recordings reaches 45%. In real-life environments like a meeting room or a noisy open space, the gain varies from 10 to 40%.