End-to-end spoofing speech detection and knowledge distillation under noisy conditions

Abstract

State-of-the-art spoofing speech detection methods highly depend on hand-crafted features like LFCC, MFCC, CQCC, FFT spectrum, etc. But almost for all the features, during extracting procedure, some information will be lost permanently, which may be of vital importance to spoofing detection task. Most speeches on the web, are often accompanied by music or noise. Using monaural speech separation to get human voice and then doing spoofing detection later may be a natural idea, but it has two shortcomings: two models must be trained, and during first separation step, some voice will be filtered out while noise will be kept. In this paper, we propose an end-to-end anti-spoofing model which fully consists of one-dimensional convolutional neural networks for spoofing detection under noisy conditions. The proposed system achieved an EER of 1.50% on the ASVspoof2019 logical access evaluation set. Under noisy conditions, we proposed a novel approach based on knowledge distillation, and our proposed method achieved great performance improvements compared to multi-condition training method.