Speech confrontation sample generation method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to a method and device for generating speech confrontation samples, electronic equipment and a storage medium. The method includes: receiving a target text, and extracting a text feature sequence from the target text; inputting the text feature sequence into a pre-trained The acoustic model of the multidimensional acoustic parameter sequence is obtained; the multidimensional acoustic parameter sequence is input into the pre-trained vocoder model to generate a time-domain sample sequence of speech as an adversarial sample corresponding to the target text, and the output of the acoustic model is a multidimensional acoustic parameter sequence, so that the generated speech content can guarantee high similarity (matching) under the description of various acoustic feature dimensions. Detection ability, more effectively deceive the speech generation detection model.

Description

technical field [0001] The present disclosure relates to the field of speech technology, and in particular, to a method and apparatus for generating a speech confrontation sample, an electronic device, and a storage medium. Background technique [0002] At present, in order to capture more discriminative information, the speech generation detection model uses a variety of acoustic features for speech signal processing, and the acoustic features for speech generation detection are directly fed into the model or used as a basis for discrimination. When generating speech adversarial samples, the speech synthesis model usually selects only one type of speech acoustic feature to model the acoustic model, and uses a vocoder to reconstruct the parameters into speech waveforms. When the acoustic parameters used by the detection model are inconsistent, because the parameters of the detection features used to generate speech are quite different from the real speech, they are easily de...

AI Technical Summary

Problems solved by technology

When generating speech adversarial samples, the speech synthesis model usually selects only one speech acoustic feature for acoustic model modeling, and uses a vocoder to reconstruct the parameters into a speech waveform. When the acoustic parameters used by the detection model are inconsistent, because the parameters of the detection features used to generate speech are quite different from the real speech, it is very easy to be detected by the speech generation detection model, and it is impossible to deceive the speech generation detection system

[0003] In addition, the existing technology mainly generates speech adversarial samples by adding random perturbation to the error threshold and clamping the error, which belongs to the passive addition of adversarial samples. Although the speech generation detection model can be deceived to a certain extent, the added noise is easy to cause The sense of hearing of the generated speech is reduced, and it is easy to be recognized and detected from the subjective perspective of humans, and this method does not start from the speech generation detection mechanism, and the generation of adversarial samples is too limited, and can only effectively deceive part of the given speech generation detection model

Images