A low-complexity, background-noise reduction preprocessor for speech encoder

An algorithm is presented for reducing stationary, though possible changing, background noise in speech time series. A perfect reconstruction filter bank decomposes the full-band time series into uniformly spaced subbands. A detection test is performed in each subband, and the results are used to dynamically adjust the gain applied to each subband. Upon reconstruction of the full-band time series, the background noise level is reduced relative to the speech. For noisy handset-speech time series, 12 to 18 dB of noise reduction can be achieved in real-world settings. As demonstrated using GSM-standard speech coders, the method can dramatically enhance the quality of coded speech