Abstract
Energy-based endpoint detection is commonly used in time domain analyses of speech segments of extracted signals to reduce the amount of computation required. However, this approach may extract incorrect speech segments due to interference by noise, which can significantly impair its recognition ability when analyzing sound files recorded in the wild. In contrast, entropy-based endpoint detection performs better in terms of noise suppression. Unfortunately, background noise that has a non-stationary frequency distribution causes drastic fluctuations in entropy values of silent segments, and weakens endpoint detection. This paper proposes using average energy entropy (AEE) endpoint detection to address these issues, and compares the AEE method with 3 other endpoint detection methods-energy-based, zero-crossing rate, and entropy-based detection methods. In experiments on frog voice-print recognition, 18 types of frog croaks recorded from the wild were analyzed, and the results revealed that the AEE method had the optimal endpoint extraction capability; and when used in concert with the linear predicative cepstral coefficients, Mel-frequency cepstrum coefficients with dynamic time warping algorithm, the AEE capability for recognition was optimized.