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

A computer-readable storage medium having encoded thereon computer-executable instructions that when executed by a processor cause the processor to perform steps comprising:

• receiving a first speech signal after it has passed though a filter;
• receiving a second speech signal;
• forming an extended feature vector by forming values for a first set of frequency components of the extended feature vector from the received first speech signal and estimating values for a second set of frequency components of the extended feature vector based on a model;
• forming a wideband feature vector by forming values for all of the frequency components of the wideband feature vector from the second speech signal;
• and
• using the extended feature vector and the wideband feature vector to train an acoustic model.

A method of training an acoustic model, the method comprising:

• collecting training speech signals over a telephone network;
• converting the training speech signals into feature vectors;
• extending the feature vectors by estimating values for frequency components that are missing in the feature vectors to form extended feature vectors;
• and
• using the extended feature vectors to train the acoustic model.