Isolated word recognition based on speech structures and discriminant analysis

Non-linguistic factors of speech such as vocal tract sizes and recording devices easily change acoustic features of speech. Recently, a new representation of speech with complete cancelation of these changes has been proposed. This representation discards the absolute properties of speech events and captures only the contrasts among them. As a full set of the contrasts in the events can define a unique geometrical structure, the proposal can be regarded as structural representation. In this paper, the new representation is examined based on two kinds of isolated word recognition tasks, a five-vowel-sequence word set and a phonetically balanced word set. Here, two problems, too strong invariance and too high dimensionality, are solved by multiple stream structuralization and linear discriminant analysis. To compare the conventional method and the proposed one, frequency-warped utterances are also used for testing. The experimental results show the high robustness of our proposed method.