Loading...

Improving the Performance of HMM-Based Speech Enhancement Techniques

Mariooryad, Soroosh | 2010

669 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 40552 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Sameti, Hossien
  7. Abstract:
  8. Various applications of speech processing systems are gaining more and more attention. Performances of these systems are dramatically affected in presence of background noise. Attempts toward improving quality of degraded speech have emerged into filed of speech enhancement which is one of the interesting topics in speech processing domain. In this thesis various single channel speech enhancement methods are studied briefly. Statistical methods as one of the most successful methods are studied more deeply. These methods are divided into two categories of short-term modeling of speech and long-term modeling. The aim of this thesis is to improve the performance of long-term methods of statistical speech enhancement systems. These approaches are recognized as model-based speech enhancement. Hidden Markov model (HMM), as the main tool in long-term methods has been studied in detail. In this work we have applied Minimum Mean Square Error (MMSE) Estimation method based on HMM to estimate clean signal in noisy environment. Since in conventional HMM-based methods, statistical models are not updated during estimation phase model mismatch is one of the drawbacks of these methods. Different pronunciations and intonations may affect speech gain. Besides, gain of noise may vary remarkably from one environment to the other. This may lead in a mismatch between energy contour of trained models and energy contour of noisy speech signal. Various methods have been proposed to tackle this problem. In this thesis we will deal with this problem with an algebraic method. In the proposed method, gain of speech and noise model will be estimated during enhancement phase. This estimation is based on matching gain of degraded signal with sum of gains of noise and speech with Least Square Error (LSE) criteria in autocorrelation domain. With this estimation noisy model is updated in each frame continuously. With more accurate models, more accurate filters are extracted and applied for enhancing degraded speech. One of drawbacks of the proposed method is the high computation time required to estimate the gains. Regarding this algebraic gain estimation mechanism, new noisy model has been proposed in speech enhancement paradigm to compensate this deficiency. Similar gain estimation mechanism is developed for the proposed method. The performance of the proposed enhancement method is evaluated using SNR and PESQ evaluation criteria. Experimental results confirm advantages of this method in presence of non-stationary noise especially in higher SNR levels
  9. Keywords:
  10. Hidden Markov Model ; Uncoherent Noise ; Minimum Mean Square Error Filters ; Speech Enhancement ; Least Squeres Error (LSE)

 Digital Object List

  • محتواي پايان نامه
  •   view

 Bookmark

No TOC