Speech enhancement has wide application areas such as speech recognition, coding, hearing aids, etc. Among various speech enhancement algorithms, Kalman filter-based ones have been known to show reliable performances in stationary noisy conditions and have advantages in DSP implementation aspects. In this thesis, we proposed a simplified version of the Kalman filter. The simplification was performed by utilizing only the diagonal components of the error covariance matrix to calculate the Kalman gain and the motivation of the simplification was also included.
Required memory units and computation load were analyzed and compared quantitatively with those of the original Kalman filter and the well-known spectral subtraction algorithm. The analysis results show that the proposed algorithm can save a lot of memory units and computation amount and thus our proposed algorithm can be easily implemented on a conventional DSP chip. We could see that our simplified Kalman filter requires only about 0.3 MIPS when the fast convergence property of the Kalman gain is also utilized.
The MMSE of our simplified algorithm was also analyzed mathematically. Although the MMSE was increased a little compared with that of the original Kalman filter, the degree of the improvement in speech quality and recognition performances for synthetic and real noises is very similar. Addition to this, our algorithm outperforms the prevailing spectral subtraction algorithm in most cases. Finally, it can be concluded that our simplified Kalman filtering algorithm can be successfully applicable to various speech application areas in adverse noisy conditions.