A data-driven method for discovering and predicting allophonic variation

Use of regression trees as a paradigm for discovering and predicting allophonic variation is reported. The current methodology is contrasted with the traditional means of allophonic rule discovery and description, namely, the hand collection of transformational rules. The tree-based regression method is developed, and an illustration of its use if provided. Specifically, results of experiments involving a database containing approximately 12000 American English stop consonants are reported. Sample rules are provided and implications for the phonological parsing component of an automatic speech recognition system are considered