A Three-Stage Cascade for Natural-Language-Driven Jazz Drum MIDI Generation: A Hybrid Keyword-Regression Control Path and Its Within-Subject Listening Evaluation

This paper reports a three-stage cascade that converts free-form natural-language descriptions into editable jazz drum MIDI sequences. The first stage parses beat-level keywords and derives a regression target vector; the second stage trains a Ridge-regression controller that maps 22 semantic dimensions onto 16 mix-ratio parameters, while a keyword-driven logic path handles on/off and fill-in triggers; the third stage renders the symbolic pattern into MIDI via a deterministic sequencer. To evaluate the system, a within-subject listening study (N = 12) compared machine-generated grooves with human-played references across four prompt types: Style-only, Style+Density, Style+Instruments, and Full description. Paired t-tests and Cliff’s delta effect sizes were calculated for each perceptual dimension. Results show no significant differences in naturalness, groove quality, or style consistency between generated and human performances for any prompt type (all p > .05). However, Bayesian estimation of the mean difference using a region of practical equivalence (ROPE) defined as ±0.3 on the 1–7 Likert scale revealed that most comparisons fell outside the ROPE, indicating that non-significance does not imply equivalence. Specifically, for Style+Instruments prompts, the naturalness difference was 0.42 [0.07, 0.77] (ROPE: ±0.3), and for Full descriptions, the groove difference was 0.44 [0.09, 0.79]. These findings demonstrate that while statistical tests fail to reject the null hypothesis, the observed effect sizes suggest the system has not yet achieved perceptual equivalence to human drummers. The study concludes that non-significance is not evidence of equivalence in the absence of an equivalence test.