Perceptually motivated speech recognition and mispronunciation detection

Abstract: This doctoral thesis is the result of a research effort performed in two fields of speech technology, i.e., speech recognition and mispronunciation detection. Although the two areas are clearly distinguishable, the proposed approaches share a common hypothesis based on psychoacoustic processing of speech signals. The conjecture implies that the human auditory periphery provides a relatively good separation of different sound classes. Hence, it is possible to use recent findings from psychoacoustic perception together with mathematical and computational tools to model the auditory sensitivities to small speech signal changes.The performance of an automatic speech recognition system strongly depends on the representation used for the front-end. If the extracted features do not include all relevant information, the performance of the classification stage is inherently suboptimal. The work described in Papers A, B and C is motivated by the fact that humans perform better at speech recognition than machines, particularly for noisy environments. The goal is to make use of knowledge of human perception in the selection and optimization of speech features for speech recognition. These papers show that maximizing the similarity of the Euclidean geometry of the features to the geometry of the perceptual domain is a powerful tool to select or optimize features. Experiments with a practical speech recognizer confirm the validity of the principle. It is also shown an approach to improve mel frequency cepstrum coefficients (MFCCs) through offline optimization. The method has three advantages: i) it is computationally inexpensive, ii) it does not use the auditory model directly, thus avoiding its computational cost, and iii) importantly, it provides better recognition performance than traditional MFCCs for both clean and noisy conditions.The second task concerns automatic pronunciation error detection. The research, described in Papers D, E and F, is motivated by the observation that almost all native speakers perceive, relatively easily, the acoustic characteristics of their own language when it is produced by speakers of the language. Small variations within a phoneme category, sometimes different for various phonemes, do not change significantly the perception of the language’s own sounds. Several methods are introduced based on similarity measures of the Euclidean space spanned by the acoustic representations of the speech signal and the Euclidean space spanned by an auditory model output, to identify the problematic phonemes for a given speaker. The methods are tested for groups of speakers from different languages and evaluated according to a theoretical linguistic study showing that they can capture many of the problematic phonemes that speakers from each language mispronounce. Finally, a listening test on the same dataset verifies the validity of these methods.