Diagnostic Evaluation of Phonetic Feature Extraction Engines: A Case Study with the Time Map Model

This paper presents a framework for evaluating phonetic feature extraction engines in a phone identification task. The case study involves HMM-based feature extraction engines for fricative and vocalic which are evaluated both at the feature level and also on how they perform when coupled with a knowledge-based phone identification model. An exact comparison model is defined and performance of the feature extraction engines is measured with respect to the degradation in accuracy as each individual feature or combination of features are introduced incrementally into the input data. This type of diagnostic evaluation facilitates a more detailed investigation of how each feature impacts on the performance of the system as a whole and provides important insights for enhancing the performance of feature extraction engines in the context of automatic speech recognition.