On the Use of Speech Recognition Techniques to Identify Bird Species

Computational Linguistics and Chinese Language Processing Vol. 19, No. 1, March 2014, pp. 55-68 55  The Association for Computational Linguistics and Chinese Language Processing On the Use of Speech Recognition Techniques to Identify Bird Species Wei-Ho Tsai and Yu-Zhi Xue Abstract Wild bird watching has become a popular leisure activity in recent years. Very often, people can see birds or hear their sounds, but have no idea what kind of bird species they are seeing. To help people learn to identify bird species from their sounds, we apply speech recognition techniques to build an automatic bird sound identification system. In this system, two acoustic cues are used for analysis, timbre and pitch. In the timbre-based analysis, Mel-Frequency Cepstral Coefficients (MFCCs) are used to characterize the bird sound. Then, we use Gaussian Mixture Models to represent the MFCCs as a set of parameters. In the pitch-based analysis, we convert bird sounds from their waveform representations into a sequence of MIDI notes. Then, Bigram models are used to capture the dynamic change information of the notes. We chose the top ten common bird species in the Taipei urban area to examine our system. Experiments conducted using audio data collected from commercial CDs and websites show that the timbre-based, pitch-based, and the combination thereof systems achieve 71.1%, 72.1%, and 75.0% accuracy of bird sound identification, respectively. Keywords: Bird Species Identification, Bigram Model, Gaussian Mixture Model, Pitch, Timbre 1. Introduction There are more than nine thousand and seven hundred bird species in the world. Although a number of birds are commonly seen, most people cannot recognize any of them. In this study, we attempt to develop automated techniques for identifying bird species from their sounds. Hereafter, this problem is referred to as bird sound identification. It is hoped that the  Department of Electronic Engineering & Graduate Institute of Computer and Communication Engineering, National Taipei University of Technology, No.1, Sec. 3, Chunghsiao E. Rd. Taipei City, 10608, Taiwan, Tel.: +886-2-27712171; Fax: +886-2-27317120 E-mail: whtsai@ntut.edu.tw The author for correspondence is Wei-Ho Tsai. 56 Wei-Ho Tsai & Yu-Zhi Xue techniques can help people learn about such animals by simply recording the bird sounds they hear and sending the recording to our system. Up to now, there has been very limited published research devoted to bird sound identification. In (Anderson et al., 1996), Anderson et al. used dynamic time warping to measure the differences in spectrogram between an unknown bird sound recording and the template bird sound recordings. In (Kogan & Margoliash, 1998), Kogan et al. compared the performance of bird sound identification obtained with dynamic time warping and hidden Markov model, in which six acoustic features were used: linear predictive coding coefficients (LPCs), LPC-derived cepstral coefficients, LPC reflection, Mel-Frequency Cepstral Coefficients (MFCCs), log mel-filter bank channel, and linear mel-filter bank channel. In (McIlraith & Card, 1997), McIlraith et al. used a backpropagation neural network and multivariate statistics to perform bird sound identification. The acoustic features tested in (McIlraith & Card, 1997) are the number of …

Content truncated to 500 words