Automatic sound-imitation word recognition from environmental sounds focusing on ambiguity problem in determining phonemes

Sound-imitation words (SIWs), or onomatopoeia, are important for computer human interactions and the automatic tagging of sound archives. The main problem in automatic SIW recognition is ambiguity in the determining phonemes, since different listener hears the same environmental sound as a different SIW even under the same situation. To solve this problem, we designed a set of new phonemes, called the basic phoneme-group set, to represent environmental sounds in addition to a set of the articulation-based phoneme-groups. Automatic SIW recognition based on Hidden Markov Model (HMM) with the basic phoneme-groups is allowed to generate plural SIWs in order to absorb ambiguities caused by listener- and situation-dependency. Listening experiments with seven subjects proved that automatic SIW recognition based on the basic phoneme-groups outperformed that based on the articulation-based phoneme-groups and that based on Japanese phonemes. The proposed system proved more adequate to use computer interactions.