The Kriston AI System for the VoxCeleb Speaker Recognition Challenge 2022

This technical report describes our system for track 1, 2 and 4 of the VoxCeleb Speaker Recognition Challenge 2022 (VoxSRC-22). By combining several ResNet variants, our submission for track 1 attained a minDCF of 0:090 with EER 1:401%. By further incorporating three fine-tuned pre-trained models, our submission for track 2 achieved a minDCF of 0:072 with EER 1:119%. For track 4, our system consisted of voice activity detection (VAD), speaker embedding extraction, agglomerative hierarchical clustering (AHC) followed by a re-clustering step based on a Bayesian hidden Markov model and overlapped speech detection and handling. Our submission for track 4 achieved a diarisation error rate (DER) of 4.86%. The submissions all ranked the 2nd places for the corresponding tracks.

Small-Footprint Keyword Spotting with Multi-Scale Temporal Convolution

Keyword Spotting (KWS) plays a vital role in human-computer interaction for smart on-device terminals and service robots. It remains challenging to achieve the trade-off between small footprint and high accuracy for KWS task. In this paper, we explore the application of multi-scale temporal modeling to the small-footprint keyword spotting task. We propose a multi-branch temporal convolution module (MTConv), a CNN block consisting of multiple temporal convolution filters with different kernel sizes, which enriches temporal feature space. Besides, taking advantage of temporal and depthwise convolution, a temporal efficient neural network (TENet) is designed for KWS system. Based on the purposed model, we replace standard temporal convolution layers with MTConvs that can be trained for better performance. While at the inference stage, the MTConv can be equivalently converted to the base convolution architecture, so that no extra parameters and computational costs are added compared to the base model. The results on Google Speech Command Dataset show that one of our models trained with MTConv performs the accuracy of 96.8% with only 100K parameters.