ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Robust endpoint detection of speech in the presence of noise

Maria Rangoussi, Stylianos Bakamidis, George Carayannis

Endpoint detection of noisy speech signals is the problem addressed in this work. Existing endpoint detection methods perform poorly on speech signal recordings from low SNR environments, thus justifying the use of more sophisticated - yet more computationally expensive - methods. Two different methods are proposed here, based on the singular value decomposition (SVD) of the autocorrelation or the observation matrix of the speech signal. Modifications that reduce the computational complexity, as well as adaptive forms are also investigated. Results and relative merits of the proposed methods as compared to existing ones are illustrated on simulated noisy speech data.