Realtime Chord Recognition of Musical Sound: a System Using Common Lisp Music — Proceedings of the International Computer Music Conference (ICMC 1999)

Claims

1. bass-discarded
   Fujishima noted that the PCP carries enough information to determine chord type and root but discards bass-note information, and proposed bass-note tracking and the use of musical context as future work.

   "It conveys information sufficient to determine the chord type and root, though it discards the bass note information."

   Locator: section: 5 · Quote language: en
2. chord-type-templates
   Fujishima matched each PCP against Chord Type Templates for 27 chord types, rotating the PCP to score every possible root and choosing the best-scoring chord.

   "I chose 27 chords (Table 1) in 2 groups, based on if a chord stays within an octave(S1) or not(S2)."

   Locator: section: 2.3 · Quote language: en
3. future-work
   Fujishima listed a better classifier, template tuning, bass-note tracking and use of musical context for error correction as future work.

   "Future work will include a better classifier, the table tuning technique, bass note tracking, use of musical context for the symbol level error correction."

   Locator: section: 5 · Quote language: en
4. keyboard-results
   On chords played with pure tones and three timbres of an electronic keyboard, the nearest-neighbour method with the initial templates distinguished all 27 chords.

   "In the nearest neighbor method, initial CTT's worked well in segregating all the 27 chords played in any of the pure tone, Occalina, Strings1, and GrandPno."

   Locator: section: 4.1 · Quote language: en
5. moldau-results
   On a 50-second excerpt from the opening theme of Smetana's Moldau, with templates tuned for the music, the system reached 94% accuracy (196 of 208 guesses); templates tuned for keyboard input did not give meaningful results on real music, and vice versa.

   "The final accuracy reached 94% (196/208). Through the experiment I found that the tables tuned for keyboard input did not give meaningful results in actual music, and vice versa."

   Locator: section: 4.2 · Quote language: en
6. pitch-class-profile
   Fujishima's system derives a Pitch Class Profile (PCP) from the spectrum: a twelve-dimensional vector of the intensities of the twelve semitone pitch classes.

   "It is a twelve dimension vector which represents the intensities of the twelve semitone pitch classes."

   Locator: section: 2.2 · Quote language: en
7. smoothing
   Assuming that a chord usually lasts for several frames, Fujishima's system averaged PCPs over time to reduce noise.

   "Assuming that a chord usually lasts for several PCP frames, this algorithm does smoothing over past PCP's by an averaging operation to reduce the noise."

   Locator: section: 2.4 · Quote language: en
8. transcription-errors
   Fujishima noted that chord recognition was traditionally approached by first transcribing individual notes and then inferring the chord symbolically, and that this approach suffers from errors in the transcription stage caused by noise and overlapping harmonics.

   "This approach suffers from recognition errors at the first stage. The errors result from noises, and from overlapping harmonics of notes in the spectrum of the input audio signal."

   Locator: page: 464 · Quote language: en