Vowel recognition based on FLAC acoutic features and subspace classifier

An approach is proposed in this paper for vowel recognition. For extracting the features of vowels, we employ the time-frequency feature extraction method (FLAC) which computes local auto-correlations on complex Fourier values. The FLAC feature takes advantage of both magnitude and phase information and extracts temporal dynamics in time and frequency domains. At recognition stage, we develop a (complex) subspace classifier to categorize the input vowels based on exploring the deviation distances from the test vowel to the trained vowel subspaces. We evaluate the proposed method by conducting experiments on a Japanese vowel dataset. The comparison experiments demonstrate the effectiveness of the proposed approach.