• The Journal of the Acoustical Society of Korea
• /
• v.37 no.6
• /
• pp.483-488
• /
• 2018

Far field input utterance is one of the major causes of performance degradation of speaker verification systems. In this study, we used teacher student learning framework to compensate for the performance degradation caused by far field utterances. Teacher student learning refers to training the student deep neural network in possible performance degradation condition using the teacher deep neural network trained without such condition. In this study, we use the teacher network trained with near distance utterances to train the student network with far distance utterances. However, through experiments, it was found that performance of near distance utterances were deteriorated. To avoid such phenomenon, we proposed techniques that use trained teacher network as initialization of student network and training the student network using both near and far field utterances. Experiments were conducted using deep neural networks that input raw waveforms of 4-channel utterances recorded in both near and far distance. Results show the equal error rate of near and far-field utterances respectively, 2.55 % / 2.8 % without teacher student learning, 9.75 % / 1.8 % for conventional teacher student learning, and 2.5 % / 2.7 % with proposed techniques.

• Journal of Convergence for Information Technology
• /
• v.7 no.6
• /
• pp.159-166
• /
• 2017

In this paper, we have experimented to improve the verification performance of speaker change detection on broadcast news. It is to enhance the input noisy speech and to apply the KL distance $D_s$ using the SNR-based weighting function $w_m$. The basic experimental system is the verification system of speaker change using GMM-UBM based KL distance D(Experiment 0). Experiment 1 applies the input noisy speech enhancement using MMSE Log-STSA. Experiment 2 applies the new KL distance $D_s$ to the system of Experiment 1. Experiments were conducted under the condition of 0% MDR in order to prevent missing information of speaker change. The FAR of Experiment 0 was 71.5%. The FAR of Experiment 1 was 67.3%, which was 4.2% higher than that of Experiment 0. The FAR of experiment 2 was 60.7%, which was 10.8% higher than that of experiment 0.

• Journal of Convergence Society for SMB
• /
• v.6 no.4
• /
• pp.71-77
• /
• 2016

In this paper, we proposed a verification of speaker change utilizing the KL distance based on GMM-UBM to improve the performance of conventional BIC based Speaker Change Detection(SCD). We have verified Conventional BIC-based SCD using KL-distance based SCD which is robust against difference of information volume than BIC-based SCD. And we have applied GMM-UBM to compensate asymmetric information volume. Conventional BIC-based SCD was composed of two steps. Step 1, to detect the Speaker Change Candidate Point(SCCP). SCCP is positive local maximum point of dissimilarity d. Step 2, to determine the Speaker Change Point(SCP). If ${\Delta}BIC$ of SCCP is positive, it decides to SCP. We examined verification of SCP using GMM-UBM based KL distance D. If the value of D on each SCP is higher than threshold, we accepted that point to the final SCP. In the experimental condition MDR(Missed Detection Rate) is 0, FAR(False Alarm Rate) when the threshold value of 0.028 has been improved to 60.7%.