ConRec 0.1 PDF Tisk Email

Czech Continuous Speech Recognition in Real Time (2002)


Within this project we have developed the first continuous speech recognition for Czech that can work with a vocabulary containing up to 20 000 most frequent words. We have used several optimization strategies, such as efficient computation of HMM probability densities, pruning schemes applied to HMM states, words and word hypotheses, a bigram compression technique as well as parallel implementation of the real recognition system. On a 2 GHz computer the system can display the recognized text in time leas than 1 s after the end of the utterance. In sentences with no OOV words the recognition rate is about 80 %.

The system is based on the one-pass strategy that processes a speech signal in time-synchronous way by efficiently combining acoustic scores of word models with language model probabilities. The words are represented by HMMs (Hidden Markov Models) that are constructed from a small set of elementary phoneme HMMs. The language model employs bigram probabilities estimated from a large text corpus. The short response time (< 1 s) has been made possible by the optimization of the search procedure, efficient bigram handling due parallel multi-thread implementation. The latter can be exploited on a PC with the MS Windows NT/2000 operating system.


More information:
  • NOUZA, J.: Strategies for Developing a Real-Time Continuous Speech Recognition System for Czech Language. In Proc. of TSD 2002. Brno, September 2002. pp. 189-196. ISBN 0302-9743