Research on Generative Artificial Intelligence Assisted Methods for Guzheng Composition

The creation of Guzheng music involves a coordinated management of modes, melody, phrasing, and performance techniques, whereas the generic symbolic music generator often represents pitch and rhythm better than specific Chinese instrumental characteristics. A framework for controlling the process of generative artificial intelligence creation based on the solo scores with the licensed or open access with REMI event encoding and the six-layer Transformer decoder is created. Modes, tempo, two-bar motif seeds, and tokens of sliding, bending, vibrato, and tremolo techniques are included in the list of generation conditions explicitly. Three configurations are tested on the basis of pitch distribution distance, rhythmic similarity, modal tone ratio, repetition frequency, novelty of n-grams, structural length validity, and evaluation by experts. Full model brings the value of pitch distribution distance to 0.098 ± 0.019, increases the modal tone ratio to 0.934 ± 0.024, and achieves the structural length validity of 0.946 ± 0.025. Expert evaluation scores become 4.47 ± 0.34 in stylistic consistency and 4.42 ± 0.37 in playability, while novelty score remains 4.16 ± 0.39. These findings show that the combination of symbolic control and conditioning with Guzheng technique brings an improvement of structural validity, stylistic appropriateness, and usability without reduction of novelty.