Flow-based prediction: a method for improved speech recognition

Most speech recognition systems are unable to cope with data from high-resolution pre-processors (such as auditory models and high-resolution spectral estimates) for two reasons. One is due to the inappropriateness of measures related to the Euclidean distance. The other is somewhat less obvious, but is due to the non-ergodic nature of short-term parameterisations of speech sounds. This aspect of speech variability is addressed. The authors show how a linear, but nonstationary, vector predictor, based on the concept of `acoustic flow´, can be used to estimate the redundancy in speech data, paving the way for an improvement in recognition performance