×

Channel-compensated low-level features for speaker recognition

  • US 10,347,256 B2
  • Filed: 09/19/2017
  • Issued: 07/09/2019
  • Est. Priority Date: 09/19/2016
  • Status: Active Grant
First Claim
Patent Images

1. A system for generating channel-compensated low level features for speaker recognition, the system comprising:

  • an acoustic channel simulator configured to receive a recognition speech signal, degrade the recognition speech signal to include characteristics of an audio channel, and output a degraded speech;

    a first feed forward convolutional neural network configured, in a training mode, to receive the degraded speech signal, and to derive from the degraded speech signal a plurality of channel-compensated low-level features, and further configured, in a test and enrollment mode, to receive the recognition speech signal and to calculate from the recognition speech signal a plurality of the channel-compensated low-level features;

    a speech signal analyzer configured, in the training mode, to extract features of the recognition speech signal;

    a loss function processor configured to calculate a loss based on the features from the speech analyzer and the channel-compensated low-level features from the first feed forward convolutional neural network;

    wherein, the calculated loss at each of a plurality of training iterations is lowered by modifying one or more connection weights of the first feed forward convolutional neural network, andif the calculated loss is less than or equal to the threshold loss, or a maximum number of training iterations has been met, the training mode is terminated.

View all claims
  • 3 Assignments
Timeline View
Assignment View
    ×
    ×