Bird Song Synthesis Based on Hidden Markov Models

Abstract

This paper focuses on the synthesis of bird songs using Hidden Markov Models (HMM). This technique has been widely used for speech modeling and synthesis. However, features and contextual factors typically used for human speech are not appropriate for modeling bird songs. Moreover, while for speech we can easily control the content of the recordings, this is not the case for bird songs, where we have to rely on the spontaneous singing of the animal. In this work we briefly overview the characteristics of bird songs, compare them to speech, and propose strategies for adapting the widely-used HTS (HMM-based Speech Synthesis System) framework to model and synthesize bird songs. In particular, we focus on Chaffinch species and a database of recordings of several song bouts of one male bird. At the end we discuss the synthesis results obtained.

Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires

“…Preliminary work has already been done in this area, for example, generating syllables of birdsong as stimuli for psychophysical and neurophysiological probes [36,37] using deep neural networks. HMMs have also been used to synthesize vocalizations [93]. With the recent advancements in machine learning, especially in areas such as generative modeling and text-to-speech, the synthesis of high-fidelity animal vocal signals is likely to become an important avenue for studying the full spectrum of vocal communication in more biologically realistic ways.…”

Bird detection in audio: A survey and a challenge

“…Also note that a sizable number of bird sounds are non-tonal, and the very rapid pitch modulation of some bird sounds can cause problems for standard frequency trackers. (The tracking methods of [20,21] account for this issue. )…”

Toward a Computational Neuroethology of Vocal Communication: From Bioacoustics to Neurophysiology, Emerging Tools and Future Directions

“…For example, the source-filter models of birdsong described above can well describe the dynamics of zebra finch song, but not the dual-syringeal dynamics of European starling song. Without reference to explicit hypotheses about underlying production mechanism, HMM based source-filter approaches provide one potential solution to this problem birdsong (Bonada et al, 2016 ).…”