Frame-level phoneme classification using inductive inference

This paper proposes a novel approach to frame-level classification by the use of inductive inference (decision trees). The proposed system (Samouelian, 1994a) uses the C4.5 induction system (Quinlan, 1993, 1996) to capture the knowledge about the structure and characteristics of the speech signal explicitly from the database. The decision tree is generated automatically from the training speech database. The database contains labelled examples in the form of a feature vector and its corresponding label for each frame. The feature vector may consist of any number of different feature sets and the label may be at the phoneme, sub-word or word level. This approach allows the integration of features from existing signal processing techniques that are currently used in stochastic modelling such as hidden Markov models (HMMs), and acoustic-phonetic features, which have been the cornerstone of traditional knowledge-based techniques. The aim of this research is to demonstrate that induction systems can provide a viable alternative automatic speech recognition technique by allowing the combination of features from any of the above feature representations to achieve optimum classification. Using C4.5, the results on five experiments are reported. The first four experiments use a small corpus of Australian English consonants (plosives, liquids and nasals) and four different feature sets, and they report on frame-level classification results for speaker-dependent and independent modes. The fifth experiment uses the TIMIT database and the mel frequency cepstral coefficient (MFCC) feature set and reports on frame-level classification results for speaker-independent experiments on the training data and test data. (C) 1997 Academic Press Limited.