A new logarithmic energy feature for endpoint detection

In this paper, we discuss the advantages and disadvantages of two usually used features: short-time linear energy and short-time log energy. Then we present a new log energy feature, which integrates the advantages of the two usual features used and overcomes their disadvantages. Compared with the short-time linear energy, it will not confuse the silence and the consonant that has small amplitude. It also avoids the problem of making the amplitude of the noise too large. So it can well reflect the difference of different parts (speech, noise and silence) of speech signal and has a good discrimination property. Combined with the Fuzzy Clustering threshold estimation method, it significantly improves the recognizer performance.