Automatic Phoneme Border Detection to Improve Speech Recognition

摘要

A comparative study of speech recognition performance among systems trained with manually labeled corpora and systems trained with semiautomatically labeled corpora is introduced. An automatic labeling system was designed to generate phoneme labels files for all words within the corpus used to train a system of automatic speech recognition. Speech recognition experiments were performed using the same corpus, first training with manually, and later with automatically generated labels. Results show that the recognition performance is better when the training of selected diccionary, is made with automatic label files than when it is made with manual label files. Not only is the automatic labeling of speech corpora faster than manual labeling, but also it is free from the subjectivity inherent in the manual segmentation performed by specialists. The performance achieved in this work is greater than 96 %.