Computer-assisted speech mimicking

摘要

Computer-assisted speech mimicking is not only interesting but is also a security threat. It involves the transformation of the speech produced by one speaker into the speech seemingly spoken by another speaker. We have built such a system that can be used to attack text prompted speaker verification systems by producing concatenated speech from authentic speech waveform segments of the impersonation target. The technique is based on the replacement of waveform segments of the impostor's speech with the corresponding waveform segments of the target speech. This paper presents two automatic speech mimicking algorithms that are based on the variable-length waveform segment replacement.