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:
obtaining, by a computer, a recognition speech signal;
generating, by the computer, a first degraded speech signal according to a first characteristic;
applying, by the computer, a neural network on the first degraded speech signal to generate a first set of low-level features;
modifying, by the computer, the first characteristic to generate a second characteristic;
generating, by the computer, a second degraded speech signal for the recognition speech signal according to the second characteristic;
applying, by the computer, the neural network on the second degraded speech signal to generate a second set of low-level features; and
generating, by the computer, a trained neural network in response to determining that a plurality of sets of low-level features generated from the neural network satisfy a loss threshold.
2. The method according to claim 1 , wherein determining that the sets of low-level features generated from the neural network satisfy the loss threshold comprises:
calculating, by the computer, a loss result according to a loss function based on the sets of low-level features.
3. The method according to claim 2 , further comprising modifying, by the computer, one or more connection weights of the neural network to lower the loss result calculated by the computer.
4. The method according to claim 1 , wherein generating the first degraded speech signal includes:
adding, by the computer, at least one environmental noise type to the recognition speech signal obtained by the computer.
5. The method according to claim 4 , further comprising selecting, by the computer, the at least one environmental noise type from a plurality of environmental noise types stored in a noise profile database.
6. The method according to claim 1 , wherein generating the first degraded speech signal includes:
simulating, by the computer, a reverberation according to a direct-to-reverberation ratio (DRR) applied to the recognition speech signal obtained by the computer.
7. The method according to claim 1 , wherein generating the first degraded speech signal includes:
simulating, by the computer, a set of one or more audio acquisition device characteristics according to an audio acquisition device profile applied to the recognition speech signal obtained by the computer.
8. The method according to claim 7 , further comprising selecting, by the computer, the audio acquisition device profile from a plurality of audio acquisition device profiles stored in a device profile database.
9. The method according to claim 1 , wherein generating the first degraded speech signal includes:
simulating, by the computer, a set of audio channel transcoding characteristics according to a transcoding profile applied to the recognition speech signal obtained by the computer.
10. The method according to claim 9 , further comprising selecting, by the computer, transcoding profile from a plurality of transcoding profiles stored in a transcoding profile database.
11. The method according to claim 1 , wherein each set of low-level features includes at least one of: Mel-frequency cep strum coefficients (MFCCs), low-frequency cepstrum coefficients (LFCCs), perceptual linear prediction (PLP) coefficients, linear or Mel filter banks, and glottal features.
12. A system comprising:
a processor; and
a non-transitory storage medium containing instructions that when executed by the processor causes the processor to perform operations comprising:
obtain a recognition speech signal;
generate a first degraded speech signal according to a first characteristic;
apply a neural network on the first degraded speech signal to generate a first set of low-level features;
modify the first characteristic to generate a second characteristic;
generate a second degraded speech signal for the recognition speech signal according to the second characteristic;
apply the neural network on the second degraded speech signal to generate a second set of low-level features; and
generate a trained neural network in response to determining that each set of low-level features generated from the neural network satisfy a loss threshold.
13. The system according to claim 12 , wherein the processor is further configured to calculate a loss result according to a loss function based on the sets of low-level features.
14. The system according to claim 13 , wherein the processor is further configured to modify one or more connection weights of the neural network to lower the loss result calculated by the processor.
15. The system according to claim 14 , further comprising a noise profile database configured to store a plurality of environmental noise types,
wherein the processor is configured to select the at least one environmental noise type from the plurality of environmental noise types stored in the noise profile database.
16. The system according to claim 12 , wherein the processor is further configured to add at least one environmental noise type to the recognition speech signal obtained by the processor.
17. The system according to claim 12 , wherein the processor is further configured to simulate a reverberation according to a direct-to-reverberation ratio (DRR) applied to the recognition speech signal obtained by the processor.
18. The system according to claim 12 , wherein the processor is further configured to simulate a set of one or more audio acquisition device characteristics according to an audio acquisition device profile applied to the recognition speech signal obtained by the processor.
19. The system according to claim 18 , further comprising a device profile database configured to store a plurality of audio acquisition device profiles, wherein the processor is configured to select the audio acquisition device profile from the plurality of audio acquisition device profiles stored in the device profile database.
20. The system according to claim 12 , wherein the processor is further configured to simulate a set of audio channel transcoding characteristics according to a transcoding profile applied to the recognition speech signal obtained by the processor.