Speaker recognition by means of a combination of linear and nonlinear predictive models
Marcos Faúndez-Zanuy · 1999
This paper presents our recent efforts in developing a speaker independent LVCSR engine for Mandarin Chinese using our multilingual database GlobalPhone. We describe a two pass approach, in which the recognition first generates Pinyin hypotheses and second transform these into Chinese character hypotheses. We show how this approach can reduce complexity and increase flexibility. We evaluate and compare different systems including different base units for speech recognition as phoneme units versus syllables. Furthermore we analyze the influence of tonal information. Our currently best system shows very promising results achieving 15.0 % character error rate.