Phoneme recognition using speech image (spectrogram)

In this paper a novel feature extraction technique based on the two-dimensional DCT (discrete cosine transform) and zigzag scanning of the spectrogram is proposed. This is in contrast to conventional approaches based on single dimension analysis such as LPC, cepstral, or FFT. As a phoneme recognition task, a series of experiments were conducted on the voice stops (`b´, `d´, `g´) of the TIMIT database uttered by 630 speakers (male and female). The extracted data form the basis for input patterns for training two types of neural networks, the semi-dynamic network (TDNN), and a static network (MLP). The highest recognition rates of 77.5 and 72.4 percent were recorded for TDNN and MLP respectively. This contrasts with results of 72 percent quoted by Hwang et al. (1992) for the same phonemes spoken by 40 females