Speaker independent speech recognition method with constrained time alignment near phoneme discriminative frame

We present constrained time alignment acoustic models based on phonetic knowledge and a speaker independent speech recognition method with the models. Japanese syllable and isolated word recognition experiments show that the models have robustness to intra- and inter-speaker varieties such as acoustic diversity. Furthermore we apply the models to a large vocabulary recognizer which works on a PC in real time. It reveals the feasibility of the models in a 5168 word recognition experiment under practical conditions such as endpoint-free matching and nonstationary noise environments.