Research On Robustness Of Speaker Verification Using Gaussian Mixture Models And System Implementation

Prevailing speaker recognition systems can obtain very high accuracy for clean speech, but their performance will degrade rapidly in noisy environments owing to the mismatch between the acoustic models and the testing speech. Therefore, noise robust technology is a crucial problem for the application of speaker recognition system in real life. The speech databases are mainly monolingual, and the performance of speaker verification system degrades dramatically because of the mismatch between mandarin and Sichuan dialect in training and test stages. Researches on Mel-Frequency Cepstrum Coefficients (MFCC) and Gaussian Mixture Models (GMM) have been made. Then based on this and the application of speech recognition on voice controlled system, the main contribution of our work in models training aiming at its stability, veracity and robustness are as following:1. In traditional model compensating methods, the white Gaussian noise is employed to simulate environmental noises to improve the robustness in the noise environment. In this paper, multi-noise multi-SNR is employed to train GMM. Experiments result shows these models are more robust than white Gaussian noise-added model.2. Due to the mismatch between mandarin and Sichuan dialect in training and test stages, the performance of speaker verification system degrades dramatically. To solve this problem, a combined Gaussian mixture model, which is trained by proportional pooling mandarin and Sichuan dialect, is presented in this paper. Experiments demonstrate that the introduced combined Gaussian mixture model is more robust for speech mismatching between mandarin and Sichuan dialect.3. A Voice Controlled Smart real time system is researched and developed using speaker verification and isolated words recognition. The speech recognition system tested by a small number of users, with high accuracy and good real-time performance, can basically fulfill the requirement of the application in office and home noise conditions with a small number of users.