Increasing Isolated Word Recognition Performance by Training Models with Reverberant Audio

Isolated Word Recognition (IWR) can be used in different applications, including home automation and car device control. These applications often take place in reverberant environments. Reverberation causes spectral distortion, harming IWR performance. We propose a multi-condition training method that uses both reverberant and non-reverberant audio to improve its generalization capabilities. Reverberant audio is obtained by applying digital sound effects to the training dataset. We used the proposed method to train an existing, baseline IWR system. Results show increased reverberation robustness in various conditions. Therefore, the proposed method poses an important contribution to voice control applications in reverberant environments.