Method and System for Generating Training Data for an Automatic Speech Recogniser

摘要

The invention describes a method and a system for generating training data (D_T) for an automatic speech recogniser ( 2 ) for operating at a particular first sampling frequency (f_H), comprising steps of deriving spectral characteristics (S_L) from audio data (D_L) sampled at a second frequency (f_L) lower than the first sampling frequency (f_H), extending the bandwidth of the spectral characteristics (S_L) by retrieving bandwidth extending information_OBE) from a codebook (6), and processing the bandwidth extended spectral characteristics (S_LE) to give the required training data (D_T). Moreover a method and a system ( 5 ) for generating a codebook ( 6 ) for extending the bandwidth of spectral characteristics (S_L) for audio data (D_L) sampled at a second sampling frequency (f_L) to spectral characteristics (S_H) for a first sampling frequency (f_H) higher than the second sampling frequency (f_L) are described.