Speech segmentation using probabilistic phonetic feature hierarchy and support vector machines

Abstract

Abstract-We propose a method that combines a probabilistic phonetic feature hierarchy with support vector machines for segmentation of continuous speech into five classes -vowel, sonorant consonant, fricative, stop and silence. We show that by using the hierarchy, only four binary classifiers are required to recognize the five classes. Due to the probabilistic nature of the hierarchy, the method overcomes the disadvantage of the traditional acoustic-phonetic methods where the error is carried down the hierarchy. In addition, the hierarchical approach allows the use of comparable amount of training data of two classes that each binary classifier is designed to discriminate. The segmentation method with 13 knowledge based parameters performs considerably better than a context-independent Hidden Markov Model (HMM) based approach that uses 39 mel-cepstrum based parameters.

Acoustic-Phonetic Approaches for Improving Segment-Based Speech Recognition for Large Vocabulary Continuous Speech

“…Many works reported cases regarding distinctive feature-based speech recognition framework, ranging from studying and proposing measurable acoustic parameters (APs) to proposing and evaluating complete distinctive feature-based speech recognition tasks in limited domains [17][18][19][20], which yielded satisfactory results. Juneja and Espy-Wilson [21][22][23] also proposed acoustic parameters (APs) for classifying speech signals into defined manner classes. They also proposed a segmentation algorithm, and complete event-based speech recognition on a limited domain, respectively.…”

“…"Speech", "Sonorant", "Syllabic" and "Continuant", which were adopted from the researches of Juneja and Espy-Wilson [21,22] [-Continuant] indicates that there is a narrow constriction blocking the air stream in the oral cavity while uttering the sound. We can combine manners of articulation into a hierarchical structure to classify phones into "broad classes" such as silence, vowels, sonorant consonants, fricatives, and stop consonants, as shown in Fig.…”

“…5. A hierarchical structure of speech manners [21,22]. Each label at the branch represents the binary value of the parent manner and its associated probability.…”

Automatic Speech Segmentation and Recognition using Class-Specific Features

“…Discovering particular section of a speech as a meaningful unit presumes recognition of that unit; on the contrary, the recognition of the unit is possible only after segmentation. While the former approach is called top-down segmentation [5], the latter approach is called the bottom-up segmentation [6]. Generally, top-down and bottom-up approaches are integrated to harness the strengths of both approaches [7], thereby increasing the performance of the system.…”

“…Thus, it employs the information from both segmentation and classification; and eventually generates appropriate segments and labels of the continuous speech signal [8]. Alternatively, sequential segmentation and recognition approach, generates segments from the acoustic cues independent of the labels, which are then fed to the classifier to identify the labels ( [6], [9]). …”