This paper describes 100 word recognition using dynamic spectral features of speech and a neural network. Spectral features of speech are obtained by a Two-Dimensional Mel-Cepstrum (TDMC). TDMC is defined as the two-dimensional Fourier transform of mel-frequency scaled logarithm spectra in the frequency and time domains. A neural network has a three-layered feedforward network and it learns automatically using a back-propagation learning algorithm. In this studyt word recognition experiments of 100 Japanese city names uttered by 7 male speakers were carried out using smoothed dynamic and average spectral features of TDMC. The number of necessary elements of TDMC was studied. Experimental results have shown that this method gives the recognition accuracy of 98.1% for speaker-dependent 100 word recognition for multi-speaker and 94.8% for speaker-independent 100 word recognition, respectively. Keyword: Dynamic Spectral Features of Speech, Neural Network.