Utilizing Tandem Features for Text-Independent Speaker Recognition on Short Utterances

This study focuses on the application of Deep Neural Networks trained to discriminate amongst senones to improve speaker recognition performance when dealing with text-independent Short Utterance Speaker Recognition (SUSR). The features derived from the said network are theorized to be more robust given that they can eliminate any unnecessary information in the final representation of the speaker. The efficacy of these features is evaluated using the test subset of the LibriSpeech speech corpus. It is found that the system's performance, especially when dealing with SUSR is greatly improved when these above-mentioned features are concatenated with traditional, more widely used Mel Frequency Cepstral Coefficients (MFCC) as measured in terms of Equal Error Rate (EER).