Method and apparatus for learning deepfake voice detection model
AssigneeFoundation of Soongsil University-Industry Cooperation
InventorsSou Hwan Jung, Ki Hun Hong, Doan Thien-Phuc
Life of this patent
Estimating the term…
Abstract
A method for learning a deepfake voice detection model and an apparatus therefor are disclosed. The method for learning a deepfake voice detection model comprises (a) resynthesizing a real voice sample using a plurality of different vocoders to generate a fake sample; (b) configuring a plurality of mini-batches using the real voice sample and the fake sample; and (c) obtaining a vocoder signature by performing supervised contrastive learning on a fake voice detection model using the plurality of mini-batches.
Claims13
A method for learning a deepfake voice detection model comprising:
- (a) resynthesizing a real voice sample using a plurality of different vocoders to generate a fake sample;
- (b) configuring a plurality of mini-batches using the real voice sample and the fake sample;
- and
- (c) obtaining a vocoder signature by performing supervised contrastive learning on a fake voice detection model using the plurality of mini-batches,
- wherein the configuring the plurality of mini-batches comprises configuring a mini-batch for a positive class,
- wherein the configuring the mini-batch for the positive class comprises,
- selecting the real voice sample as an anchor sample;
- selecting two real voice samples different from the anchor sample as a positive sample;
- selecting an augmented sample that augments the anchor sample;
- and
- selecting a plurality of fake samples generated corresponding to the anchor sample as a negative sample.
A computing apparatus comprising:
- a memory storing at least one instruction;
- and
- a processor executing the instruction stored in the memory,
- wherein each instruction executed by the processor performs steps comprising,
- (a) resynthesizing a real voice sample using a plurality of different vocoders to generate a fake sample;
- (b) configuring a plurality of mini-batches using the real voice sample and the fake sample;
- and
- (c) obtaining a vocoder signature by performing supervised contrastive learning on a fake voice detection model using the plurality of mini-batches,
- wherein the configuring the plurality of mini-batches comprises configuring a mini-batch for a positive class,
- wherein the configuring the mini-batch for the positive class comprises,
- selecting the real voice sample as an anchor sample;
- selecting two real voice samples different from the anchor sample as a positive sample;
- selecting an augmented sample that augments the anchor sample;
- and
- selecting a plurality of fake samples generated corresponding to the anchor sample as a negative sample.
Dependencies are inferred from the claim text.