In a knowledge-based speech recognition system, landmarks are key points in time in the speech waveform. They guide the search for the underlying distinctive features. The effect of background noise on the automatic detection of landmarks is studied. Two landmark detection algorithms are tested. The full algorithm is hierarchical and uses many criteria to look for speech cues. The minimalist algorithm uses a reduced set of these criteria. Experiments show that the full algorithm outperforms the minimalist algorithm in clean speech, while the minimalist algorithm outperforms the full algorithm in very noisy speech. These results suggest that, in the presence of background noise, a speech recognition system should customize its algorithm and parameter values to the characteristics of the noise in order to perform effectively.
Bibliographic reference. Liu, Sharlene A. (1995): "Noise effects on landmark detection in a speech recognition system", In EUROSPEECH-1995, 1567-1570.