Training wideband acoustic models in the cepstral domain using mixed-bandwidth training data for speech recognition

Abstract

A method and apparatus are provided that use narrowband data and wideband data to train a wideband acoustic model.

1 . A method of training an acoustic model, the method comprising:
 using values in a first set of training vectors that represent all of the frequency components in a set of frequency components and using values in a second set of training vectors that represent fewer than all of the frequency components in the set of frequency components to train a set of spectral domain acoustic model parameters; and   converting the set of spectral domain acoustic model parameters into a set of cepstral domain acoustic model parameters.   
   
   
       2 . The method of  claim 1  further comprising converting the set of cepstral domain acoustic model parameters into a past set of spectral domain acoustic model parameters and using the past set of spectral domain acoustic model parameters to train a new set of spectral domain acoustic model parameters. 
   
   
       3 . The method of  claim 2  wherein converting the cepstral domain acoustic model parameters into a past set of spectral domain acoustic model parameters comprises converting a set of cepstral domain acoustic means into a past set of spectral domain acoustic means by applying an inverse transform to each cepstral domain acoustic mean to form an inverse transformed cepstral domain mean and adding a same constant value to each inverse transformed cepstral domain mean. 
   
   
       4 . The method of  claim 3  wherein the set of cepstral domain acoustic means comprises means for a plurality of mixture components. 
   
   
       5 . The method of  claim 2  wherein converting the cepstral domain acoustic model parameters into a past set of spectral domain acoustic model parameters comprises converting a set of cepstral domain acoustic covariances into a past set of spectral domain acoustic covariances by applying inverse transforms to each cepstral domain covariance to form an inverse transformed covariance and adding a same constant value to each inverse transformed covariance. 
   
   
       6 . The method of  claim 1  wherein using values that represent fewer than all of the frequency components comprises not using values of a selected frequency component in any of the second set of training vectors. 
   
   
       7 . The method of  claim 1  wherein training a set of spectral domain acoustic model parameters comprises identifying a conditional mean for a frequency component that is not represented by the values used from the second set of training vectors. 
   
   
       8 . The method of  claim 7  wherein identifying a conditional mean comprises identifying a separate conditional mean for each training vector in the second set of training vectors. 
   
   
       9 . The method of  claim 8  wherein identifying a conditional mean for a training vector in the second set of training vectors comprises identifying the conditional mean based on the values of the training vector. 
   
   
       10 . The method of  claim 1  wherein the cepstral domain model parameters comprise Hidden Markov Model parameters. 
   
   
       11 . A computer storage medium having computer-executable instructions for performing steps comprising:
 setting values for model parameters for a Hidden Markov Model in the cepstral domain;   converting the cepstral model parameters to the spectral domain to form spectral model parameters;   modifying the spectral model parameters; and   converting the spectral model parameters to the cepstral domain.   
   
   
       12 . The computer storage medium of  claim 11  wherein converting the model parameters to the cepstral domain comprises applying a truncated transform to the model parameters. 
   
   
       13 . A computer storage medium having computer-executable instructions for performing steps comprising:
 converting cepstral domain acoustic model means into a past set of spectral domain acoustic model means by applying an inverse transform to each cepstral domain acoustic mean to form an inverse transformed cepstral domain mean and adding a same constant value to each inverse transformed cepstral domain mean;   using values in a first set of training vectors that represent all frequency components in a set of frequency components and using values in a second set of training vectors that represent fewer than all of the frequency components in the set of frequency components together with the past set of spectral domain acoustic model means to train a set of spectral domain acoustic model parameters; and   converting the set of spectral domain acoustic model parameters into a set of cepstral domain acoustic model parameters.