Redução de ruído para sistemas de reconhecimento de voz utilizando subespaços vetoriais.

Abstract

The establishment of a speech-based communication interface between humans and computers has been pursued since the beginning of the computer era. Several studies have been made over the last six decades in order to accomplish this interface, making possible commercial use of speech recognition applications. However, factors such as noise, reverberation, distortion among others degrades the performance of these systems. Thus, reducing their success rate when operating in adverse environments. With this in mind, the study of techniques to reduce the impact of these problems is of a great value and has gained prominence in recent decades. The work presented in this dissertation aims to reduce problems related to noise encountered in an automotive environment, improving the speech recognition system robustness. Thus,controlofnon-critical features of a car, such as CD player and air conditioning, can be performed through voice commands. The proposed system is based on a speech signal preprocessing step using the signal subspace method. Its performance is related to the size (lines× columns) of the matrices that represents the input signal. Therefore, the ULLV decomposition was used because it offers a lower computational complexity compared to traditional methods based on SVD decomposition. The speech recognizer Julius is an open source software that offers high performance and was the chosen one for the case study. A noisy speech database with 44800 samples was generated to model the automotive environment. Finally, the robustness of the system was evaluated and compared with a traditional method of noise reduction called spectral subtraction.