Proc. V International Conference on Advances in Pattern Recognition (ICAPR 2003)
Hidden Markov models (HMM) are widely used in speech recognition these days. For all practical pu... more Hidden Markov models (HMM) are widely used in speech recognition these days. For all practical purposes, viz, mathematical tractability and reducing computational complexity, the observation density function in a continuous HMM is assumed to be mixture gaussian with diagonal covariance matrices. The feature transformation plays an important role in achieving this decorrelation. Different feature transformations like discrete cosine transform (DCT) and principal component analysis (PCA) have been tried. DCT is a suboptimal transform that nearly diagonalizes the covariance matrix and retains higher order correlations. PCA on the other hand diagonalizes the covariance matrix but again retains the higher order dependencies. Independent component analysis (ICA), used widely in blind source separation problems, can be used as a feature transform.
ICA makes the feature vectors independent and thereby fully decorrelated. Experiments are conducted on a 50-word vocabulary in a multi-speaker mode. The results are compared with the baseline system where DCT is used as a feature transform. A 2% increase in recognition accuracy is observed using ICA as a feature transformation over DCT.
Proc. V International Conference on Advances in Pattern Recognition (ICAPR 2003)
Hidden Markov models (HMM) are widely used in speech recognition these days. For all practical pu... more Hidden Markov models (HMM) are widely used in speech recognition these days. For all practical purposes, viz, mathematical tractability and reducing computational complexity, the observation density function in a continuous HMM is assumed to be mixture gaussian with diagonal covariance matrices. The feature transformation plays an important role in achieving this decorrelation. Different feature transformations like discrete cosine transform (DCT) and principal component analysis (PCA) have been tried. DCT is a suboptimal transform that nearly diagonalizes the covariance matrix and retains higher order correlations. PCA on the other hand diagonalizes the covariance matrix but again retains the higher order dependencies. Independent component analysis (ICA), used widely in blind source separation problems, can be used as a feature transform.
ICA makes the feature vectors independent and thereby fully decorrelated. Experiments are conducted on a 50-word vocabulary in a multi-speaker mode. The results are compared with the baseline system where DCT is used as a feature transform. A 2% increase in recognition accuracy is observed using ICA as a feature transformation over DCT.
Uploads
Papers by Joel Pinto
ICA makes the feature vectors independent and thereby fully decorrelated. Experiments are conducted on a 50-word vocabulary in a multi-speaker mode. The results are compared with the baseline system where DCT is used as a feature transform. A 2% increase in recognition accuracy is observed using ICA as a feature transformation over DCT.
ICA makes the feature vectors independent and thereby fully decorrelated. Experiments are conducted on a 50-word vocabulary in a multi-speaker mode. The results are compared with the baseline system where DCT is used as a feature transform. A 2% increase in recognition accuracy is observed using ICA as a feature transformation over DCT.