Goal

• To transcribe solo-voice melody automatically.
• The recording is of Taiwanese “cantillation” (yínsòng 吟誦): a classical poem is read in “reading pronunciation” (different from spoken language) and the tones of the words are elaborated artistically into a melody.
• The poem in question is the “Ballad of the Lute” (pípá xíng 琵琶行) of Bái Jūyì 白居易 (772–846), Ms. Hsü I-t‘ing 許禕娗 singing, c. 2009. The melody is that taught by Mr. Hong Zenan 洪澤南.

First listen to the opening eight lines (of eighty-eight):

First attempt

• After many experiments, I settled on WaoN) (和音 ‘harmonious notes’), an open-source command-line tool for converting convert .wav files into MIDI.
• The WaoN output heard here was converted to MP3 via FluidSynth) using the “GeneralUser” SoundFont (lines 1-2 only). GeneralUser models the sound of the piano, and I judge it less unpleasant than most other SoundFonts for analytical purposes.

First attempt

• Result:

• Terrible! Tremolo in the voice and reverb in the recording lead to chords or trills in MIDI transcription. Stray notes abound.

Second attempt

• For comparison, I also paid a professional transcriber to produce a transcription (here, lines 1-2 only).

• Professional transcription:

• Is this significantly better than the WaoN version? (The paid transcriber relied on Finale 2011 and professional training.)
• It seems to me that the WaoN version is truer to the feeling of the original, though the paid transcriber’s is better for having fewer notes.

New idea

• Can we clean MIDI with Python?
• There are MIDI modules for Python, but I felt a purer approach would be to convert MIDI to text (using MIDICSV) and write a script to discard all but the notes with highest “note-on velocity” (playback volume and intensity). I used a priority queue to hold notes and selected the highest-velocity note at each time-tick.

• Result:

  Still terrible!

Finally

• Eventually I learned to use MuseScore, an open source program for music transcription, with export to MIDI, MXL (MusicXML), and many other formats.
• The human brain seems to have won this round. I learned a lot by making the attempt, though.
• Below is the MuseScore transcription and play-back (first eight lines of the poem).

I'd be grateful for any corrections you may find to make in my transcription.

END