End-to-end speaker recognition using deep neural network
The present invention is directed to a deep neural network (DNN) having a triplet network architecture, which is suitable to perform speaker recognition. In particular, the DNN includes three feed-forward neural networks, which are trained according to a batch process utilizing a cohort set of negative training samples. After each batch of training samples is processed, the DNN may be trained according to a loss function, e.g., utilizing a cosine measure of similarity between respective samples, along with positive and negative margins, to provide a robust representation of voiceprints.
1. A computer-implemented method comprising:
receiving, by a computer, a recognition speech sample;
executing, by the computer, a neural network on the recognition speech sample to generate a voiceprint of the recognition speech sample, the neural network being a portion of a triplet neural architecture trained using dual sets of positive speech samples and a cohort set of negative speech samples;
comparing, by the computer, the generated voiceprint with at least one stored voiceprint; and
performing, by the computer, a speaker recognition on the recognition speech sample based upon the comparison.
2. The computer-implemented method of claim 1 , wherein performing the speaker recognition comprises:
comparing, by the computer, the generated voiceprint with the at least one stored voiceprint, wherein the at least one stored voiceprint is associated with a speaker to be verified.
3. The computer-implemented method of claim 1 , wherein the performing the speaker recognition comprises:
comparing, by the computer, the generated voiceprint with the at least one stored voiceprint, wherein the at least one stored voiceprint is associated with a closed set of known speakers.
4. The computer-implemented method of claim 3 , wherein the closed set of known speakers is a blacklist associated with telephone fraud.
5. The computer-implemented method of claim 1 , further comprising:
pre-processing, by the computer, the recognition speech sample prior to executing the neural network.
6. The computer-implemented method of claim 5 , wherein the preprocessing of the recognition speech sample comprises:
segmenting, by the computer, the recognition speech sample into windows of a predetermined duration with a predetermined window shift; and
extracting, by the computer, features to be fed into the neural network from each window.
7. The computer-implemented method of claim 1 , wherein performing the speaker recognition comprises at least one of:
identifying, by the computer, a speaker associated with the recognition speech sample; and
verifying, by the computer, the speaker associated with the recognition speech sample.
8. A system comprising:
a non-transitory storage medium storing a plurality of computer program instructions; and
a processor electrically coupled to the non-transitory storage medium and configured to execute the plurality of computer program instructions to:
receive a recognition speech sample;
deploy a neural network on the recognition speech sample to generate a voiceprint of the recognition speech sample, the neural network being a portion of a triplet neural architecture trained using dual sets of positive speech samples and a cohort set of negative speech samples;
compare the generated voiceprint with at least one stored voiceprint; and
perform a speaker recognition on the recognition speech sample based upon the comparison.
9. The system of claim 8 , wherein the processor is configured to further execute the computer program instructions to:
compare the generated voiceprint with the at least one stored voiceprint, wherein the at least one stored voiceprint is associated with a speaker to be verified.
10. The system of claim 8 , wherein the processor is configured to further execute the computer program instructions to:
compare the generated voiceprint with the at least one stored voiceprint, wherein the at least one stored voiceprint is associated with a closed set of known speakers.
11. The system of claim 10 , wherein the closed set of known speakers is a blacklist associated with telephone fraud.
12. The system of claim 8 , wherein the processor is configured to further execute the computer program instructions to:
pre-process the recognition speech sample prior to executing the neural network.
13. The system of claim 12 , wherein to pre-process the recognition speech sample, the processor is configured to further execute the computer program instructions to:
segment the recognition speech sample into windows of a predetermined duration with a predetermined window shift; and
extract features to be fed into the neural network from each window.
14. The system of claim 8 , wherein the speaker verification includes at least one of the processor identifying a speaker associated with the recognition speech sample and the processor verifying the speaker associated with the recognition speech sample.
15. A computer-implemented method comprising:
feeding, by a computer, a first set of one or more positive speech samples attributed to a speaker to a first feed-forward neural network to generate a first embedding vector;
feeding, by the computer, a second set of one or more positive speech samples attributed to the speaker to a second feed-forward neural network to generate a second embedding vector;
feeding, by the computer, a cohort set of negative speech samples not attributed to the speaker to a third feed-forward neural network to generate a set of embedding vectors;
calculating, by the computer, a loss function based upon the first embedding vector, the second embedding vector, and set of embedding vectors; and
back-propagating, by the computer, the loss function to modify one or more connection weights in each of the first, second, and third feed-forward neural networks.
16. The computer-implemented method of claim 15 , wherein the loss function is based upon a positive distance corresponding to a degree of similarity between the first and second embedding vectors and a negative distance corresponding to a degree of similarity between the first embedding vector and an embedding vector most similar to the first embedding vector in the set of embedding vectors.
17. The computer-implemented method of claim 16 , wherein each of the degree of similarity between the first and second embedding vectors and the degree of similarity between the first embedding vector and the embedding vector most similar to the first embedding vector in the set of embedding vectors is based upon a cosine measure.
18. The computer-implemented method of claim 15 , wherein the loss function is based upon an equal error rate metric associated with a degree of similarity between the first and second embedding vectors and a degree of similarity between the first embedding vector and an embedding vector most similar to the first embedding vector in the set of embedding vectors.
19. The computer-implemented method of claim 15 , further comprising:
initializing, by the computer, at least one of the first, second, and third feed-forward neural network with random connection weights.
20. The computer-implemented method of claim 15 , further comprising:
initializing, by the computer, at least one of the first, second, and third feed-forward neural network with connection weights of a deep neural network trained to perform speaker identification on a closed set of speakers.