Phonetic segmentation using statistical correction and multi-resolution fusion

Abstract

This paper focuses on the generation of accurate phonetic segmentations. Statistical methods based on absolute and relative correction are discussed and experimented on both monophone and biphone models to improve the segmentation results. The influence of search range on the statistical correction process is studied and a state selection technique is used to enhance the correction results. This paper also explores the influence of resolution (stepsize) of HMMs and proposes a multi-resolution fusion process to further refine the statistically corrected results. Improvements of segmentation results in terms of segmentation accuracy, mean absolute error (MAE), and root mean square error (RMSE) can be observed by applying the proposed refinement methods.

A hybrid refinement scheme for intra- and cross- corpora phonetic segmentation

“…All the proposed refinement steps contribute to the improvements of segmentation results in terms of accuracy and MAE/RMSE. Compared with the previously reported work in Zhao et al (2013), the results presented in this paper are improved due to the use of isolated-unit training to obtain improved baseline models as suggested in Donovan (1996), and Yuan et al (2013), the inclusion of both CI and CD models in the fusion process, and the application of predictive models for refinements. As presented in Table 4, the proposed scheme exhibits higher accuracies on TIMIT as compared with recent studies.…”

“…As the automatically detected boundary is defined by the onset of the first state of a phone, it is possible to calculate the correction term as a ratio, i.e., a relative term, of the state-level segmentations around the automatically detected boundaries (Zhao et al, 2013):…”

“…The scheme presented in this paper is an extension of the previous work reported in Zhao et al (2013). It is a hybrid refinement scheme for phonetic segmentation consisting of three components: (1) statistical correction which addresses systematic biases of acoustic models using local information from the most relevant range (i.e., "state-selection" in Section 3); (2) a fusion method which incorporates complementary effects from acoustic models with various resolutions (i.e., stepsizes -the interval to extract each frame of feature vector) which are demonstrated to affect the results significantly; and (3) predictive models which correct non-systematic segmentation biases using predictive models.…”