A probabilistic framework for feature-based speech recognition

Abstract

Most current speech recognizers use an observation space which is based on a temporal sequence of "frames" (e.g., Mel-cepstra). There is another class of recognizer which further processes these frames to produce a segment-based network, and represents each segment by fixed-dimensional "features." In such feature-based recognizers the observation space takes the form of a temporal network of feature vectors, so that a single segmentation of an utterance will use a subset of all possible feature vectors. In this work we examine a maximum a posteriori decoding strategy for feature-based recognizers and develop a normalization criterion useful for a segmentbased Viterbi or A search. We report experimental results for the task of phonetic recognition on the TIMIT corpus where we achieved context-independent and context-dependent (using diphones) results on the core test set of 64.1% and 69.5% respectively.

Application of Kernel-Based Feature Space Transformations and Learning Methods to Phoneme Classification

“…The more sophisticated segmental techniques fit parametric curves to the feature trajectories of the phonemes [24]. There is, however, a much simpler methodology [13] that applies non-uniform smoothing and sampling in order to parametrize any phoneme with the same number of features, independent of its length. it allows us to apply any sort of machine learning algorithm to the phoneme classification task.…”

Characterizing and Processing Robot-Directed Speech

“…Extracting OOV phone sequences Recognizer is that developed by the SLS group at MIT [8]. The recognizer used the OOV model developed by Bazzi in [3].…”

Using aggregation to improve the performance of mixture Gaussian acoustic models

“…Phonetic recognition accuracies were computed on both a 50 speaker development test set (400 utts) and on the 24 speaker core test (192 utts). The recognizer used for these experiments is described in detail in [3].…”

“…Each experiment utilizes the SUMMIT [3] recognition system. This system uses mixture Gaussian acoustic models to score segment-based phonetic units.…”