Short-time Fourier Inversion Based Multi-band Generation for Efficient Synthetic Speech Attack

Recent research indicates that automatic speech recognition (ASR) systems are highly susceptible to adversarial examples. Current ASR attack methods can be divided into two categories: audio-dependent and audio-independent attacks, based on whether they rely on benign audio inputs. Among them, the Speech Synthesizing based Attack (SSA), a typical audio-independent method, has gained significant attention due to its flexibility and generalizability. However, SSA suffers from low attack efficiency owing to the complex structure of the VITS decoder. To alleviate this problem, we propose a novel ASR attack method called Short-time Fourier Inversion-based Multi-band Generation (SFI-MG), which offers two main advantages. Firstly, we employ an improved inverse short-time Fourier transform (iSTFT) to optimize the traditional residual blocks, accelerating the upsampling process. Secondly, we introduce a pseudo-orthogonal mirror filter bank module to enable parallel processing across multiple frequency bands, thereby enhancing both processing speed and speech quality. Comprehensive experiments on the Common Voice, Librispeech, and VCTK datasets demonstrate that our adversarial examples not only improve attack efficiency by 210%, but also achieve a high success rate in targeted attacks.