Deep Neural Yodelling

Yodel music differs from most other genres by exercising the transition from chest voice to falsetto with an audible glottal stop which is recognised even by laymen. Yodel often consists of a yodeller with a choir accompaniment. In Switzerland, it is differentiated between the natural yodel and yodel songs. Today's approaches to music generation with machine learning algorithms are based on neural networks, which are best described by stacked layers of neurons which are connected with neurons of the lower layer. The networks are trained using real music samples in lead sheet or audio waveform to adapt their weights, so they produce similar ones. So far, proposed procedures have used piano music or polyphonic music that consisted of a classical orchestra. Approaches to vocal music were rarely found, and measured in sound quality less successful than for instrumental music.

To the best of the author's knowledge, up until this thesis, it has not been tried to produce yodel music with machine learning algorithms. In total 17h of natural yodel and 17h of yodel songs were available. For this work, convolution neural networks were used to investigate if generative approaches can produce yodel-like music in the audio waveform. Experiments with the model WaveNet which was proposed for music synthesis in the audio waveform confirmed the ability of generative methods to produce the timbre of yodel music. Two experts affirmed the conclusion in this thesis. Further investigations revealed that the model reproduces yodel-like vocalisation where high notes are sung with an "u" and breast notes with an "o" which is a typical characteristics of natural yodel. The model fails to provide a melody structure, and changes in metre are missed utterly. In summary, a WaveNet-based model successfully produced the timbre of yodel but was unable to compose yodel music. The generative approach Variational Autoencoder was investigated to tackle the shortcomings of WaveNet. Unfortunately, it did not succeed in producing anything different than noise which is explained by its short experimental time. Therefore, it is inconclusive whether the approach is appropriate for yodel music.

To conclude, the model WaveNet achieved to produce the timbre of yodel music. However, more research is necessary to generate entire yodel compositions.