While tempo fluctuation in classical music performances has been well investigated, previous score-based studies offer limited practical help to instrumental players for daily self-assessment. We introduce a score-independent, unsupervised probabilistic framework that enables musicians to situate their renditions within corpora of peer recordings. This approach respects the fundamentally subjective nature of music by eliminating the need for “correct” reference labels. In our experiments, piano and flute recordings are represented as chroma- and timbre-based feature vectors and modeled with a left-to-right Hidden Markov Model. Hidden states correspond to a virtual score that governs the musical content, such as pitch and dynamics, modeled by a Gaussian distribution, while state transition probabilities model tempo fluctuations. Viterbi decoding analyzes performer-specific trajectories of hidden states, and clustering these trajectories exhibits each player's rendition relative to peer performances. Our research suggests that our framework provides players with concrete, comparative feedback for artistic reflection and evaluation. As the method does not use symbolic scores or manual annotation, it makes performance analysis possible when scores are not available. It enables musicians to refine, compare, and enhance their interpretive choices using audio recordings of their performances. As the method does not use symbolic scores or manual annotation, it makes performance analysis possible when scores are not available.
Hiraiwa et al. (Wed,) studied this question.