The overarching aim of the proposed work is to align a promising treatment lead - Musical Intervention (MI) - with a promising mechanistic account of psychosis - Predictive Processing. This protocol focuses on the R33 phase, to optimize its administration (Is active participation more effective than passive listening? Does creation of new music help more than performing others' creations?). By tracking the interrelation between symptom mechanisms and MI, the investigators can use those metrics to prospectively assign patients to particular MI. The R33 phase will examine the impact of SING on computational behavioral metrics of (Aim 1) Conditioned Hallucinations, (Aim 2) Social Reinforcement Learning, (Aim 3) Language Use, in 200 participants with voice hearing in the context of a psychotic illness (n=50per per group). Following a screening visit to determine eligibility, these computerized tasks will be administered behaviorally, and an interview will elicit speech, prior to and following the full SING intervention (in 10 groups of 5 participants, each facilitated by a trained musical interventionist, during the first two years of the project). Participants will complete these tasks prior to and following randomization to four different conditions (facilitated by a SING team member) that will deconstruct the possible active ingredients of SING along two dimensions: Activity and Ownership: (a) SING (n=50, Activity + and Ownership +), participants produce and perform their own song; (b) Karaoke (n=50, Activity + and Ownership -), participants perform karaoke, singing along to others music; (c) Pop Music (n=50, Activity - and Ownership -), participants will listen to popular music chosen by the music interventionists; and (d) Curated Playlists (n=50, Activity -, Ownership +), participants will curate playlists of popular music and listen to them together. This deconstruction will provide insights into the predictive processing framework, as applied to hallucinations and music, specifically, whether changes at higher, a-modal, hierarchical levels, particularly sense of self and active inference, influence precision weighted perceptual and social inferences more so than inactive experiences or experiences that do not engage sense of self. This R33 portion of the study was originally included in NCT05537428, which now has results posted for the R61 phase of the study.