A trend estimation algorithm for singing pitch detection in musical recordings

Abstract

Detecting pitch values for singing voice in the presence of music accompaniment is challenging but useful for many applications. We propose a trend estimation algorithm to detect the pitch ranges of a singing voice in each time frame. The detected trend substantially reduces the difficulty of singing pitch detection by reducing a large number of wrong pitch candidates either produced by musical instruments or the overtones of the singing voice. The proposed algorithm can be applied to improve the performance of singing pitch detection. Quantitative evaluations show that proposed trend estimation improves an existing algorithm significantly. The results from the MIREX 2010 competition show that our system achieves the best overall raw-pitch accuracy for vocal songs.

An Acoustic-Phonetic Model of F0 Likelihood for Vocal Melody Extraction

“…Note that represents the Euclidean distance between the observed partial-amplitude vector and the th example partial-amplitude vector, which is converted to a similarity score by (the monotonically decreasing portion of) the Gaussian function in (5). A high value of would always imply that is the true vocal F0 because its partials exhibit vocal timbre.…”

“…Dressler [4] devised an approach where the predominant melody is tracked by an auditory streaming model that favors unstable, high-magnitude F0 contours. Hsu et al [5] extracted the relative extents of vibrato and tremolo from each partial as features for classifying vocal and instrumental partials, and implemented F0 continuity by determining a sequence of tight ranges for the vocal F0. Tachibana et al [6] used instability in F0 and intensity as well as shortness in duration to enhance melodic components.…”

“…Ellis and Poliner [9] took a purely data-driven approach, letting the difference between melodic and non-melodic F0s be learned by support vector machine from a labeled set of training data. Hsu et al [5] collected data for vocal and instrumental partials and trained 2 Gaussian mixture models for the 2-way classification. Joo et al [8] modeled the timbre of melodic source by a set of partial amplitude examples, which is derived from -means clustering of partial amplitude data points extracted from monophonic audio data.…”

“…In melodic component enhancement, Tachibana et al [6] processed instability and shortness by applying a harmonic/percussive separation algorithm to the audio in 2 passes with different window lengths. The first pass of this enhancement technique was adopted as a preprocessing step in [5].…”

“…In our implementation, acoustic-phonetic F0 modification expands a small 84-sample set of vocal timbre examples by a factor of 88. In contrast, the existing data-driven approaches [5], [8], [9] apply a supervised or unsupervised learning procedure to a large audio data set composed of vocal and instrumental sound sources.…”