A study of recognition rate improvement for Thai esophageal speech by using feature conversion

The paper represents a front-end process for changing esophageal speech features into normal speech features in order to improve a recognition rate of esophageal speech in a speech recognition system that training by normal speech corpus based on Hidden Markov Models (HMMs). A system, that combines feature conversion technique and cepstral normalization technique in order to prevent variation bias in each signal, is proposed. From experimental results, the feature conversion technique can significantly improve the recognition rate of esophageal speech from 33.75% up to 67.48%. Moreover, the recognition rate is raised up to 77.68% when using the cepstral normalization technique with the feature conversion.