Speech Bandwidth Expansion For Speaker Recognition On Telephony Audio

Practical applications often require speaker recognition systems to work well for audio files of different sampling rates. However, the performance of speaker recognition systems degrades substantially under a mismatched audio sampling rate between the training and testing conditions. For example, wideband speaker recognition models trained on audio files with a 16kHz sampling rate perform poorly on telephony audio with an 8kHz sampling rate due to the missing higher frequency information. In this paper, we propose a Deep Neural Network (DNN) based system to estimate the speech spectrum in the frequencies above 4kHz for narrowband 8kHz telephony audio. We train the proposed system on speech datasets processed using various simulated telephony codecs. Additionally, we perform speaker recognition experiments by using the bandwidth expansion system as a preprocessor for speaker verification using wideband models. The evaluation datasets used for speaker verification are codec-degraded downsampled Voxceleb1 and SITW, and the NIST SRE 2010 10s-10s condition. We see a significant improvement in the results compared to a simple upsampling with interpolation and low-pass filtering. Additionally, these promising experiments show that the proposed bandwidth expansion system can be used successfully as a data augmentation for training speaker embedding systems.