Can Large-Scale Vocoded Spoofed Data Improve Speech Spoofing Countermeasure with a Self-Supervised Front End?

A speech spoofing countermeasure (CM) that discriminates between unseen spoofed and bona fide data requires diverse training data. While many datasets use spoofed data generated by speech synthesis systems, it was recently found that data vocoded by neural vocoders were also effective as the spoofed training data. Since many neural vocoders are fast in building and generation, this study used multiple neural vocoders and created more than 9,000 hours of vocoded data on the basis of the VoxCeleb2 corpus. This study investigates how this large-scale vocoded data can improve spoofing countermeasures that use data-hungry self-supervised learning (SSL) models. Experiments demonstrated that the overall CM performance on multiple test sets improved when using features extracted by an SSL model continually trained on the vocoded data. Further improvement was observed when using a new SSL distilled from the two SSLs before and after the continual training. The CM with the distilled SSL outperformed the previous best model on challenging unseen test sets, including the ASVspoof 2019 logical access, WaveFake, and In-the-Wild.

1. INTRODUCTION

The detection of spoofed speech generated by text-to-speech (TTS) and voice conversion (VC) systems is usually formulated as a binary classification task [1]. A detector, referred to as a spoofing countermeasure (CM), requires a significant amount of training data containing diverse human (bona fide) and synthesized (spoofed) speech waveforms. However, preparing various spoofed training data is costly. For example, it took a few months to develop the TTS and VC systems that generated the training set of the ASVspoof 2019 logical access (LA) database through trial and error [2].