作者: H. Hild , A. Waibel
DOI: 10.1109/ICASSP.1993.319284
关键词: Pattern recognition 、 Network architecture 、 Entropy (information theory) 、 Artificial intelligence 、 Speech recognition 、 Hidden Markov model 、 Time delay neural network 、 Multi state 、 Artificial neural network 、 Unsupervised learning 、 Computer science
摘要: The authors present an improved multistate time delay neural network (MS-TDNN) for speaker-independent, connected letter recognition which outperforms HMM (hidden Markov model) based system (SPHINX) and previous MS-TDNNs. They also explore new architectures with internal speaker models. Four different characterized by increasing number of speaker-specific parameters are introduced. can be adjusted automatic identification or adaptation, allowing tuning-in to a speaker. Both methods lead significant improvements over the straightforward speaker-independent architecture. Even unsupervised (speech is unlabeled) works well. >