Deep Learning Approaches for Sung Vowel Classification

Phoneme classification is an important part of automatic speech recognition systems. However, attempting to classify phonemes during singing has been significantly less studied. In this work, we investigate sung vowel classification, a subset of the phoneme classification problem. Many prior approaches that attempt to classify spoken or sung vowels rely upon spectral feature extraction, such as formants or Melfrequency cepstral coefficients. We explore classifying sung vowels with deep neural networks trained directly on raw audio. Using VocalSet, a singing voice dataset performed by professional singers, we compare three neural models and two spectral models for classifying five sung Italian vowels performed in a variety of vocal techniques. We find that our neural models achieved accuracies between 68.4% and 79.6%, whereas our spectral models failed to discern vowels. Of the neural models, we find that a fine-tuned transformer performed the strongest; however, a convolutional or recurrent model may provide satisfactory results in resource-limited scenarios. This result implies that neural approaches trained directly on raw audio, without extracting spectral features, are viable approaches for singing phoneme classification and deserve further exploration.