Datasets

- Jamendo, 93 songs, annotations of presence/absence of singing on fractions of second scale
- singing detection as a task - has been around since the beginning of MIR, has not been solved in general;
- how should silence within a word be interpreted? very soft singing? onsets are not unambiguous
- singing detection vs vocal phrases, Brendan’s topic of interest
- other datasets mentioned in relation to singing detection task: all source separation datasets which have stems - about 100 songs taken together; Cantometrics tracks which are almost all singing tracks, Polina also produced frame wise annotations for about 100 tracks; polyphonyproject.com - all singing tracks; EPFL Montreaux Jazz Festival have loads of recordings with stems, if Simon gets access