| Abstract [eng] |
Nowadays speech processing should become very important because it is one of the best alternatives of present control means like keyboard and mouse. So this work studies: • The use of speech signal processing. • Various methods and algorithms of speech recognition. • The implementation of Lithuanian speech recognition in this work. • Comparison with other algorithms. • Simulation of voice control. There are described and analyzed these algorithms that are used in speech processing: Dynamic Time Warping [1], Linear Predictive Coding [2], Linear Spectral Pair (or Frequencies) ([2], [3], [8]), Vector Quantization [4], Adaptive Quantization [5] and Hidden Markov Models [9]. In the implementation of Lithuanian speech recognition were used Linear Predictive Coding, Linear Spectral Pair, Vector Quantization and Dynamic Time Warping algorithms, Java programming language, microphone and WAVE format. The speech recognition system were tested with more noisy and less noisy microphones, with various WAVE files recorded in different conditions and with real time speaking. Further for simulation of voice control were modeled three situations: calculator with four operations (sum, subtraction, multiplication and division), changing background colors and moving mouse cursor. To conclude it should be noted that: • Implemented speech recognition algorithm mostly gives around 90% recognition. • The precision of testing depends on microphone and surrounding noise, length and acoustic likeness of recognizable words. • Word finding and Dynamic Time Warping algorithms are optimized for most cases, but surely they can be improved for better results. |
