The algorithm we propose has, for speech input, a convergence speed four times that of the normalized LMS (NLMS) but its computational load is almost the same. This algorithm uses a different step size for each coefficient of an adaptive transversal filter, and these step sizes are time-invariant and weighted in proportion to the expected variation of a room impulse response. They are weighted this way because this expected variation becomes progressively smaller (along the series) by the same exponential with which the energy of the impulse response decays. This algorithm also reflects the whitening of the input signal (that is, it removes the correlation between input vectors) and is therefore especially for speech, whose spectrum is highly nonwhite. To reduce the computational complexity, this algorithm incorporates a fast projection routine modified for a practical DSP structure. The computational load is thus only 2L multiply-add operation. The algorithm is implemented in an echo canceller constructed with multiple DSP chips, and its fast convergence is demonstrated.
|Number of pages||8|
|Journal||NTT R and D|
|Publication status||Published - 1995|
ASJC Scopus subject areas
- Electrical and Electronic Engineering