This paper presents a method for automatic music transcription applied to audio recordings of a cappella performances with multiple singers. We propose a system for multi-pitch detection and voice assignment that integrates an acoustic and a music language model. The acoustic model performs spectrogram decomposition, extending probabilistic latent component analysis (PLCA) using a six-dimensional dictionary with pre-extracted log-spectral templates. The music language model performs voice separation and assignment using hidden Markov models that apply musicological assumptions. By integrating the two models, the system is able to detect multiple concurrent pitches in polyphonic vocal music and assign each detected pitch to a specific voice type such as soprano, alto, tenor or bass (SATB). We compare our system against multiple baselines, achieving state-of-the-art results for both multi-pitch detection and voice assignment on a dataset of Bach chorales and another of barbershop quartets. We also present an additional evaluation of our system using varied pitch tolerance levels to investigate its performance at 20-cent pitch resolution.

Harmony constitutes an essential aspect of a broad range of styles in Western music, and chords usually play a key role therein. Consequently, the generation or detection of chords is central to a wide range of computational models, for instance in chord estimation, chord sequence prediction, and harmonic structure detection. Such models are typically evaluated by comparing their outputs to ground-truth chord labels using a binary metric (“correct” or “incorrect”). As chord vocabularies continue to grow, binary metrics capture less information about the correctness of a given label, thus equating all labeling errors regardless of their severity. In this work, we present the chord-eval toolkit, which proposes three different metrics drawn, adapted, and generalized from previous work, addressing acoustic, perceptual, music-theoretical, and mechanical aspects of evaluation. We discuss use cases for which the metrics vary in appropriateness, depending on properties of the underlying music and the task at hand, and present an example of such differences.