wav to midi conversion
bibliography
Wav to midi conversion

Wav to midi converter is an algorithm that takes a sound signal as its input and attempts to capture pitch, loudness, dynamics, and performance style (vibrato, portamento, glissando) present in that sound signal, and to write them down to a MIDI file. The resulting MIDI file sound very close to original performance but it cannot be represented as sequence of notes in a standard music notation.

Music transcription

Music transcription is the act of listening to a piece of music and of writing down music notation for the piece. This is a complicated cognitive task that is performed by high-skilled human musicians.

Music transcriber is an algorithm that takes a sound signal as its input and attempts to recognize pitches, onset times, and durations of the notes that make up that signal. That is, it attempts to find the symbolic representation of music that a human listener would write down for the same input signal (assuming that there is one such representation). Music transcriber does not attempt to reproduce a perceptually close "copy" of the original performance as wav to midi converter does. In this sense, automatic music transcription is similar to the task of speech recognition. Automatic music transcription still remains unsolved scientific problem.

Transcription accuracy

There are no music transcription software giving 100% recognition accuracy for unrestricted solo performances. Solo Explorer does some errors as well. Nevertheless, we are the only ones that use quantitative measures and constantly monitor the transcription accuracy of our transcription engine. The transcription accuracy achieved by Solo Explorer v1.0 is estimated to be as follows:

  Solo Explorer v1.0
Pitch extraction (octave) error, % 0.33
Split notes, % 4.60
Missed notes, % 18.50
Rhythm recognition error, % 25.70
Tone recognition error, % 6.60

The estimates above are obtained on our internal benchmark database. This database consists of 100 folk songs (4 hours of records, over 27000 notes). The results you will obtain with Solo Explorer may be much better, as our benchmark songs are heavily loaded with grace-notes and appogiaturas. These are short music events difficult to perceive and recognize but counted to be of the same importance as notes.

     to top   

 

     Copyright © 2002 Gailius Raskinis. All rights reserved.