Low Snr Speech Recognition Technology

Speech recognition at low SNR(signal-to-noise ratio) environment is currently one of the most important research area in the world. It is also a focus issue and difficulty in speech recognition field, which has especially important theoretical and practical significance. This thesis begins with research on basic theroies of speech processing, especially emphasizes on the speech recognition algorithms, noise estimation and speech enhancement techniques, and applies speech enhancement techniques to speech recognition system in noisy environment. Experiments show that the performance of system has been improved obviously. Following is the main work of this thesis:1. After researching on the basic theory of speech recognition, this thesis realizes Talker-independent recognizer and Speaker identification system with DTW(Dynamic time warping) and VQ(Vector quantization) algorithm respectively. Improvement has been proposed to provide Speaker identification system a better performance. The experiments show that the recognition rate is upper to 96 percent in quiet environment, the recognition rate declines sharply in low SNR environment, and with certain low input SNR level, nearly unrecognizable.2. On the basis of analyzing noise property, this thesis emphasizes on two noise estimation methods: classical method based on Voice Activity Detection(VAD) and the latest method based on optimal smoothing and minimum statistics. Some improvements have been proposed to enhance the VAD's performance based on LPC Cepstrum coefficients. Experiment results show that both two methods have a good estimation to the steady noise, and the latter one has a better performance for unsteady noise.3. A variety of speech enchancement algorithms are discussed including Wiena filter, Spectral Subtraction, MMSE algorithm and the masking model combined with the Spectral Subtraction algorithm. With the combination of noise estimation and the speech enhancement algorithms, this thesis did lots of experiments at low SNR environment and analyzed the simulation results.4. Speech enhancement techniques are applied to the speech recognition system in this thesis. The experiment results show that after the combination of the two systems, the output SNR of the noise speech is improved and the recognition rate is enhanced...