EVALUATION OF A NOISE-REDUCTION METHOD - COMPARISON BETWEEN OBSERVED SCORES AND SCORES PREDICTED FROM STI

Speech was mixed with different noise signals and then processed accor ding to the well-known noise reduction method of 'spectral subtraction '. Three different algorithms were examined. The speech signals were s ubjected to a four alternative forced choice (4AFC) test. Both the pro cessed and unprocessed signals were evaluated psyche-acoustically and objectively. Speech intelligibility was measured with the 4AFC test by presenting the signals via headphones to a group of normal-hearing an d to a group of hearing-impaired listeners. The intelligibility scores were compared with the intelligibility scores predicted from a modifi ed version of the Speech Transmission Index (STI). It appeared that al though the noise reduction algorithms reduced the noise level, they di d not improve the measured speech intelligibility, either for normal-h eating or for hearing-impaired listeners. This, however, was inconsist ent with the scores estimated from STI, which erroneously predicted a significant improvement in intelligibility due to the noise reduction processing.