Self-supervised speech representations for fake audio detection
A method for determining synthetic speech includes receiving audio data characterizing speech in audio data obtained by a user device. The method also includes generating, using a trained self-supervised model, a plurality of audio features vectors each representative of audio features of a portion of the audio data. The method also includes generating, using a shallow discriminator model, a score indicating a presence of synthetic speech in the audio data based on the corresponding audio features of each audio feature vector of the plurality of audio feature vectors. The method also includes determining whether the score satisfies a synthetic speech detection threshold. When the score satisfies the synthetic speech detection threshold, the method includes determining that the speech in the audio data obtained by the user device comprises synthetic speech.
1. A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:
receiving audio data characterizing speech obtained by a user device;
generating, using a shallow discriminator model, a score indicating a presence of synthetic speech in the audio data; and
based on determining that the score satisfies a synthetic speech detection threshold, determining that the speech in the audio data obtained by the user device comprises synthetic speech,
wherein the shallow discriminator model is trained on a plurality of mixed training samples, each particular training sample of the plurality of mixed training samples comprising;
a corresponding human-originated portion being first human-originated speech representing a first portion of a corresponding training utterance; and
a corresponding synthetic speech portion being synthetic speech representing a second portion of the corresponding training utterance, the corresponding synthetic speech portion replacing second human-originated speech that represents the second portion of the corresponding training utterance for the particular training sample.
2. The computer-implemented method of claim 1 , wherein the shallow discriminator model comprises an intelligent pooling layer.
3. The method of claim 2 , wherein the intelligent pooling layer is configured to aggregate a plurality of embeddings of the audio data.
4. The computer-implemented method of claim 2 , wherein the operations further comprise:
generating, using the intelligent pooling layer of the shallow discriminator model, a single final audio feature vector based on the audio data,
wherein generating the score indicating the presence of the synthetic speech in the audio data is based on the single final audio feature vector.
5. The computer-implemented method of claim 4 , wherein the shallow discriminator model further comprises a single fully-connected layer.
6. The computer-implemented method of claim 5 , wherein the single fully-connected layer is configured to receive, as input, the single final audio feature vector and generate, as output, the score.
7. The computer-implemented method of claim 1 , wherein the shallow discriminator model comprises a logistic regression model.
8. The computer-implemented method of claim 1 , wherein the shallow discriminator model comprises a linear discriminant analysis model.
9. The computer-implemented method of claim 1 , wherein the shallow discriminator model comprises a random forest model.
10. The computer-implemented method of claim 1 , wherein the data processing hardware resides on the user device.
11. The computer-implemented method of claim 1 , wherein the shallow discriminator model comprises a single intelligent pooling layer and a single fully-connected layer.
12. A system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware, the memory hardware storing instructions that, when executed on the data processing hardware, cause the data processing hardware to perform operations comprising:
receiving audio data characterizing speech obtained by a user device;
generating, using a shallow discriminator model, a score indicating a presence of synthetic speech in the audio data; and
based on determining that the score satisfies a synthetic speech detection threshold, determining that the speech in the audio data obtained by the user device comprises synthetic speech,
wherein the shallow discriminator model is trained on a plurality of mixed training samples, each particular training sample of the plurality of mixed training samples comprising:
a corresponding human-originated portion being first human-originated speech representing a first portion of a corresponding training utterance; and
a corresponding synthetic speech portion being synthetic speech representing a second portion of the corresponding training utterance, the corresponding synthetic speech portion replacing second human-originated speech that represents the second portion of the corresponding training utterance for the particular training sample.
13. The system of claim 12 , wherein the shallow discriminator model comprises an intelligent pooling layer.
14. The system of claim 13 , wherein the intelligent pooling layer is configured to aggregate a plurality of embeddings of the audio data.
15. The system of claim 13 , wherein the operations further comprise:
generating, using the intelligent pooling layer of the shallow discriminator model, a single final audio feature vector based on the audio data,
wherein generating the score indicating the presence of the synthetic speech in the audio data is based on the single final audio feature vector.
16. The system of claim 15 , wherein the shallow discriminator model further comprises a single fully-connected layer.
17. The system of claim 16 , wherein the single fully-connected layer is configured to receive, as input, the single final audio feature vector and generate, as output, the score.
18. The system of claim 12 , wherein the shallow discriminator model comprises a logistic regression model.
19. The system of claim 12 , wherein the shallow discriminator model comprises a linear discriminant analysis model.
20. The system of claim 12 , wherein the shallow discriminator model comprises a random forest model.
21. The system of claim 12 , wherein the data processing hardware resides on the user device.
22. The system of claim 12 , wherein the shallow discriminator model comprises a single intelligent pooling layer and a single fully-connected layer.