Channel-compensated low-level features for speaker recognition
A system for generating channel-compensated features of a speech signal includes a channel noise simulator that degrades the speech signal, a feed forward convolutional neural network (CNN) that generates channel-compensated features of the degraded speech signal, and a loss function that computes a difference between the channel-compensated features and handcrafted features for the same raw speech signal. Each loss result may be used to update connection weights of the CNN until a predetermined threshold loss is satisfied, and the CNN may be used as a front-end for a deep neural network (DNN) for speaker recognition/verification. The DNN may include convolutional layers, a bottleneck features layer, multiple fully-connected layers and an output layer. The bottleneck features may be used to update connection weights of the convolutional layers, and dropout may be applied to the convolutional layers.
1. A computer-implemented method comprising:
receiving, by a computer, a speech signal containing an utterance;
extracting, by the computer, low-level channel compensated features from the speech signal by executing a convolutional neural network (CNN),
the CNN being trained to minimize a loss function between low-level channel compensated features calculated from a plurality of computer-degraded training speech signals and low-level handcrafted features from corresponding clean training speech signals; and
executing, by the computer, a deep neural network (DNN) on the extracted low-level channel compensated features to recognize a speaker of the utterance in the speech signal.
2. The computer-implemented method of claim 1 , wherein the low-level channel compensated features extracted from the speech signal include at least one of: Mel-frequency cepstrum coefficients (MFCCs), low-frequency cepstrum coefficients (LFCCs), perceptual linear prediction (PLP) coefficients, linear or Mel filter banks, and glottal features.
3. The computer-implemented method of claim 1 , wherein the DNN includes a plurality of convolutional layers, the method further comprising:
executing, by the computer, the plurality of convolutional layers on the low-level channel compensated features extracted by the CNN to increase inter-speaker variability and decrease intra-speaker variability.
4. The computer-implemented method of claim 1 , wherein the DNN includes a plurality of fully connected layers, the method further comprising:
executing, by the computer, the fully connected layers on the low-level channel compensated features extracted by the CNN and other handcrafted or learned features to increase inter-speaker variability and decrease intra-speaker variability.
5. The computer-implemented method of claim 4 , wherein one or more of the handcrafted or learned features include at least one of: MFCC, LFCCs, PLP, filter-banks, and glottal features.
6. The computer-implemented method of claim 1 , wherein the computer-degraded training speech signals include the corresponding clean training speech signals degraded with at least one of: environmental noise, reverberation, acquisition device audio artifacts, and transcoding noise.
7. The computer-implemented method of claim 1 , wherein recognizing the speaker of the utterance includes identifying the speaker.
8. The computer-implemented method of claim 7 , wherein identifying the speaker comprises:
generating, by the computer, a voiceprint of the speaker based upon executing the DNN on the extracted low-level channel compensated features; and
matching, by the computer, the generated voiceprint to at least one voiceprint in a predefined list of a plurality of voiceprints.
9. The computer-implemented method of claim 1 , wherein recognizing the speaker of the utterance includes verifying the speaker.
10. The computer-implemented method of claim 9 , wherein verifying the speaker comprises:
generating, by the computer, a voiceprint of the speaker based upon executing the DNN on the extracted low-level channel compensated features; and
comparing, by the computer, the generated voiceprint of the speaker with at least one registered voiceprint of the speaker.
11. 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 computer program instructions to:
receive a speech signal containing an utterance;
extract low-level channel compensated features from the speech signal by deploying a convolutional neural network (CNN),
the CNN being trained to minimize a loss function between low-level channel compensated features calculated from a plurality of computer-degraded training speech signals and low-level handcrafted features from corresponding clean training speech signals; and
deploy a deep neural network (DNN) on the extracted low-level channel compensated features to recognize a speaker of the utterance in the speech signal.
12. The system of claim 11 , wherein the low-level channel compensated features extracted from the speech signal include at least one of: Mel-frequency cepstrum coefficients (MFCCs), low-frequency cepstrum coefficients (LFCCs), perceptual linear prediction (PLP) coefficients, linear or Mel filter banks, and glottal features.
13. The system of claim 11 , wherein the DNN includes a plurality of convolutional layers, and wherein the processor is configured to further execute the plurality of computer program instructions to:
deploy the plurality of convolutional layers on the low-level channel compensated features extracted by the CNN to increase inter-speaker variability and decrease intra-speaker variability.
14. The system of claim 11 , wherein the DNN includes a plurality of fully connected layers, and wherein the processor is configured to further execute the plurality of computer program instructions to:
deploy the plurality of fully connected layers on the low-level channel compensated features extracted by the CNN and other handcrafted or learned features to increase inter-speaker variability and decrease intra-speaker variability.
15. The system of claim 14 , wherein one or more of the handcrafted or learned features include at least one of: MFCC, LFCCs, PLP, filter-banks, and glottal features.
16. The system of claim 11 , wherein the computer-degraded training speech signals include the corresponding clean training speech signals degraded with at least one of: environmental noise, reverberation, acquisition device audio artifacts, and transcoding noise.
17. The system of claim 11 , wherein recognizing the speaker of the utterance includes identifying the speaker.
18. The system of claim 17 , wherein the processor is configured to further execute the computer program instructions to:
generate a voiceprint of the speaker based upon deploying the DNN on the extracted low-level channel compensated features; and
match the generated voiceprint to at least one voiceprint in a predefined list of a plurality of voiceprints to identify the speaker.
19. The system of claim 11 , wherein recognizing the speaker of the utterance includes verifying the speaker.
20. The system of claim 19 , wherein the processor is configured to further execute the computer program instructions to:
generate a voiceprint of the speaker based upon deploying the DNN on the extracted low-level channel compensated features; and
compare the generated voiceprint of the speaker with at least one registered voiceprint of the speaker to identify the speaker.