Automatic Song-Type Classification and Speaker Identification of Norwegian Ortolan Bunting (Emberiza Hortulana) Vocalizations

Abstract

This paper presents an approach to song-type classification and speaker identification of Norwegian Ortolan Bunting (Emberiza Hortulana) vocalizations using traditional human speech processing methods. Hidden Markov Models (HMMs) are used for both tasks, with features including Mel-Frequency Cepstral Coefficients (MFCCs), log energy, and delta (velocity) and delta-delta (acceleration) coefficients. Vocalizations were tested using leave-one-out cross-validation. Classification accuracy for 5 song-types is 92.4%, dropping to 63.6% as the number and similarity of the songs increases. Song-type dependent speaker identification rates peak at 98.7%, with typical accuracies of 80-95% and a low end at 76.2% as the number of speakers increases. These experiments fit into a larger framework of research working towards methods for acoustic censusing of endangered species populations and more automated bioacoustic analysis methods.

Call recognition and individual identification of fish vocalizations based on automatic speech recognition: An example with the Lusitanian toadfish

“…Indeed the HMM system based on the feature set in Table III presented an average identification rate of 95.5%. This method reached values at least similar to or even higher than the identification rates observed in automatic recognition approaches with mammals' (Campbell et al, 2000;Clemins, 2005;Reby et al, 1997Reby et al, , 2006 and birds' (Terry and McGregor, 2002;Trawicki, 2005) vocalizations. The comparison, however, is not straightforward.…”

“…Indeed the identification score of these sound types was too low (<10%) to consider automatic monitoring. Some heterogeneity in the recognition rates of different sound types of a species is commonly reported (Chesmore and Ohya, 2004;Jahns, 2008;Kogan and Margoliash, 1998;Parsons and Jones, 2000;Sch€ on et al, 2001;Trawicki, 2005). Several reasons may be responsible for the low identification rate of sounds (Young et al, 2006).…”

“…The comparison, however, is not straightforward. These investigations were based on recordings involving a smaller number of animals singing simultaneously; this generally improves recognition (as exemplified in Trawicki, 2005). In addition, the acquisition of very good quality nondegraded sounds also improves recognition, such as the work carried out by Clemins et al (2005), who recorded captive elephants with microphones held by collars.…”

Multichannel speech recognition using distributed microphone signal fusion strategies

“…Based on results of previous work with song-type classification on single-channel Ortolan Bunting vocalizations [14], the analysis conditions for the distributed microphone corpus were frames of 5 ms with 50% overlap with 12 Generalized Cepstral Coefficient (GFCC) [15] features computed from the 26-channel filterbanks [16] and appended with the delta and delta-delta coefficients. The left-to-right song-type Hidden Markov Models (HMMs) [17] consisted of 18-states with a single diagonalcovariance Gaussian Mixture Model (GMM) underlying each state with approximately an equal split of the four song-types across each of the 8-channel microphones under matched training and testing conditions.…”

Generalized Perceptual Features for Vocalization Analysis Across Multiple Species

“…Song-type classification and speaker identification experiments [14] were performed on the Ortolan Bunting dataset. MFCCs, GFCCs, and GPLP-derived CCs.…”