Frog Sound Identification System Based On Automatic Syllables Segmentation Haryati Jaafar Intelligent Biometric Research Group (IBG), School of Electrical and Electronic Universiti Sains Malaysia, Engineeing Campuss 14300 Nibong Tebal, Pulau Pinang, Malaysia haryati_jaafar@yahoo.com Dzati Athiar Ramli Intelligent Biometric Research Group (IBG), School of Electrical and Electronic Universiti Sains Malaysia, Engineering Campuss 14300 Nibong Tebal, Pulau Pinang, Malaysia dzati@eng.usm.my Abstract— Automatic recognition of frog sound according to particular species is considered a worthy tool for biological research and environmental monitoring. In addition, the synthesis of peptides with antimicrobial activity found in the skin of certain frog species is valuable for medical values. As a result, automatic recognition of frog sound offers many advantages rather than manual method that depending on physical observation procedure. This study evaluate the accuracy of frog sound identification from 10 species that recorded from Malaysia forest located at Sungai Sedim, Kulim and Lata Mengkuang, Baling. By applying frequency information technique, the frog sound samples are automatically segmented into syllables. Two types of syllable feature extraction method i.e, MFCC and LPC are the determined. Finally, nonparametric kNN classifier with Euclidean distance and Chebyshev distance has been employed to recognize the frog species. Results show that kNN classifier based on MFCC and LPC are capable to identify the frog species with an accuracy of 88.75% and 95%, respectively. Keywords- Frog identification; kNN classifier; Euclidean distance; Chebyshev distance; MFCC and LPC. I. INTRODUCTION Generally, the animals generate sound either deliberately as a means of communication or as a by-product of eating, flying or locomotion. Hence, identifying animal species from their sounds is increasing interest and has been employed for application such as biological research and environmental monitoring [1]. Acoustic signal plays an important role in frog biology and has been widely studied to address questions in taxonomy ecology, and conservation. However, manual practical implementation of animal sound identification regularly heads into the problem of time consuming, logistically difficult, costly and lack of expertise to detrimental the identification of collecting sound samples. Therefore the development of a software system for pattern recognition and classification is required to meet this challenge [2]. The acoustic signals of animal sound representing as a sequence of syllables. Thus, syllables can be used as the acoustic component to identify the animal species. Different techniques of automatic segmentation have been practiced in an audio signal. The two oldest and most common speech segmentation techniques, energy [3,4] and zero crossing [5], because of their simplicity and ease of use. Some study also combines both of energy and zero-crossing [6]. However, both of segmentation methods cannot extract the full syllable due to the weakness to noise. In recent years, the recognition of frog sounds has become crucial since frog plays as important role in ecological environments. Frog is considered to be bio indicators because the health of frog population indicated the health of whole ecosystem due to their biphasic life. Also, frogs are in general susceptible to environmental toxicants since their permeable skin may provide antimicrobial peptides that useful for medical purpose [7-9]. Several techniques have been studied for frog identification to classify their species and most of these approaches rely on time-frequency domain analyses [10- 14]. However, the results from the above experiment are based on clean data which is obtained from websites where the samples are mostly recorded in the laboratory. However, the effect background noise often seriously interferes in real implementation. In the real condition, the recordings may contain interference background noise i.e., the sound of running water, and the variable nature of heather conditions i.e., the sound of the wind or other animal calls in the background caused the limitations in acquiring a clean data. This paper develops an automatic system for frog identification that basically consists of three parts namely, syllable segmentation, feature extraction and classification as shown in Fig. 1. The frog sound was recorded from locations around Baling and Kulim, Kedah uses Sony Stereo IC Recorder ICD-AX412F supported with Sony electret condenser microphone in 32-bit wav files at a sampling frequency of 48kHz. The sounds were recorded next to a