Fig 1.
Representation of music data and its conversion.
Fig 2.
Example of melody in midi file.
(Use FL Studio DAW for presentation).
Fig 3.
System flow chart.
Fig 4.
Flow chart of data preprocessing.
Fig 5.
Example of textual conversion, the graph of the upper part is the musical melody, and the bottom is the corresponding textual melody text.
Fig 6.
GPT-2 model flow chart.
Fig 7.
GPT-2 model fine-tuning.
Fig 8.
Block diagram of music evaluation model.
Fig 9.
Sample music melody text after preprocessing, and each line represents the melody of a song.
Fig 10.
The change of loss value during training, the abscissa is rounds and the ordinate is loss.
Fig 11.
Examples of using the MT-GPT-2 model to generate music, the note in red box is the starting note of the input, the notes framed in blue in (a) (b) are rhythmic pitches similar to the starting notes, while the notes framed in green in (c) (d) are rhythmic and pitch pitches unrelated to the starting notes.
Fig 12.
MIDI files, (a) and (b) generated using the LSTM model were compared with MIDI files (c) and (d) generated using the MT-GPT-2 model.
(Use FL Studio DAW for presentation).
Table 1.
Mathematical evaluation table.
Fig 13.
Smooth-saltatory progression comparison analysis.
Fig 14.
Wave test analysis.
Table 2.
Music theory evaluation table.