Quality improvement of low-bit-rate noisy speech using the subspace method

Additive noise presents a particularly difficult problem for linear predictive-based speech coding (LPC) systems operating at sampling rates up to 4.8 kb/s. The performance. of cascade signal-subspace-based speech enhancement algorithm/code excited linear predictive (CELP) coding system is studied and compared with the conventional cascade spectral-subtraction-based speech enhancement algorithm/CELP system. Both systems have been tested under the same conditions, i.e. the same type of noise (white noise) and the same signal-to-noise-ratios. An LPC-based objective test is used for the evaluation of both systems. Based on this test, it is shown that the signal-subspace-based speech enhancement algorithm outperforms the spectral-subtraction-based speech enhancement algorithm such that it could be recommended as a preprocessor for any LPC speech coding system.