Learning front-end speech recognition parameters within neural network training

Abstract

Techniques for learning front-end speech recognition parameters as part of training a neural network classifier include obtaining an input speech signal, and applying front-end speech recognition parameters to extract features from the input speech signal. The extracted features may be fed through a neural network to obtain an output classification for the input speech signal, and an error measure may be computed for the output classification through comparison of the output classification with a known target classification. Back propagation may be applied to adjust one or more of the front-end parameters as one or more layers of the neural network, based on the error measure.

What is claimed is: 
     
         1 . (canceled) 
     
     
         2 . An apparatus comprising:
 at least one processor; and   at least one storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry out a method, the method comprising:   recognizing input speech using an automatic speech recognition (ASR) engine, wherein recognizing the input speech comprises:
 processing a frame of the input speech to produce a power spectrum for the frame; 
 providing the power spectrum as input to a plurality of filters of a filter bank layer inside of a neural network of a speech recognition engine, wherein each of at least some of the plurality of filters comprises a respective set of non-negative frequency weights for a corresponding set of frequencies, and wherein the filter bank layer performs vocal tract length normalization (VTLN) frequency pooling. 
   
     
     
         3 . The apparatus of  claim 2 ,
 wherein processing the frame of the input speech to produce a power spectrum for the frame comprises producing a normalized power spectrum for the frame, and   wherein providing the power spectrum as input to the plurality of filters of the filter bank layer comprises providing the normalized power spectrum as input to the plurality of filters.   
     
     
         4 . The apparatus of  claim 2 , wherein producing the normalized power spectrum for the frame comprises:
 performing a non-linear transformation on the power spectrum for the frame to produce a non-linear power spectrum; and   normalizing the non-linear power spectrum to produce a normalized non-linear power spectrum.   
     
     
         5 . The apparatus of  claim 2 , wherein recognizing the input speech comprises performing a non-linear transformation on an output of the plurality of filters of the filter bank layer. 
     
     
         6 . The apparatus of  claim 2 , wherein the filter bank layer is a convolutional layer in the neural network of the speech recognition engine, the convolutional layer having weight sharing among the filters. 
     
     
         7 . The apparatus of  claim 2 , wherein at least some filters of the plurality of filters have multiple frequency peaks. 
     
     
         8 . The apparatus of  claim 2 , wherein each of the plurality of filters is associated with a frequency band centered at a center frequency identified by vocal-tract-length normalization filters.