 |
 | Home | Contact Us | FAQ | Search & Site Map | Link to Us |
 | Sign In | Join | Other 45 Sites in Network |
Natural Science Forum / Physics / Acoustics / May 2007Midi To Audio Questions
Hi, I have a wave recoding of a guitar playing single notes which I'd like to convert to MIDI format. I'm using a C++ program to process this audio. My first question is, how do I detext the attack points accurately? At the moment, I take a fft of short length (256 for 44.1kHz audio), and I add the values in all the bins and dividie this value by the largest value in a single bin. This lets me know how noisy or chromatic a single is at a point in time. A guitar attack is noisy so this information helps. But is there a better way? How about other instrument playing single notes such as a flute? In this case, I could uses the fft to establish when a note is on or off by summing the values in an fft window and comparing this value with some threshold. My second question is how might I calculate the loudness of this attack, in sones perhaps (although since I am working with a wave file were talking about relative loudness)? Thirdly, what would be an intelligent way of mapping the loudness of a note against its MIDI velocity. Thanks very much for your help, Barry. Barry, > I have a wave recoding of a guitar playing single notes which I'd like to > convert to MIDI format. I'm using a C++ program to process this audio. I applaud your ambition because what you are attempting is VERY difficult if not impossible. There are programs that attempt to transcribe audio into MIDI, but the ones I've seen are pretty lame. > My first question is, how do I detext the attack points accurately? Some DAW programs like Sonar and ACID look for note attacks to identify tempo and where the beats fall by watching for a sudden rise in level. This is probably more direct than using an FFT to watch for "impure" frequencies (pick noise, flute chiffs, etc). But the guitar player better be mighty clean, because even a hint of a second note will throw a wrench into the works! > How about other instrument playing single notes such as a flute? Now you're getting to the "impossible" part. If the flute player plays with a silence between each note you can look for the silence. But real music is not played that way. Real players play staccato AND legato. So now you have to watch for changes in pitch only. > My second question is how might I calculate the loudness of this attack, > in sones perhaps (although since I am working with a wave file were > talking about relative loudness)? Thirdly, what would be an intelligent > way of mapping the loudness of a note against its MIDI velocity. There's no way to map signal level in a Wave file to the original volume in the room as SPL. You'll need to decide what dB levels in the file relate to various note-on velocities. --Ethan > Hi, > [quoted text clipped - 19 lines] > note against its MIDI velocity. > Sorry to rain on your parade here, but what your attempting to do is pretty much impossible - at least with the technology available on his planet... Let me explain why starting from a Fourier transform of the music will lead you down a dead end street. Lets say you have a page of text, and you analyse it by counting how many times each character occurs. So you'll end up 138 A's, 18 B'c, 47 C's etc. Now you attempt to use the frequency count of each letter to reconstruct the original text on the page. This is really what you're trying to do with Audio -> MIDI conversion. Chris W  SignatureThe voice of ignorance speaks loud and long, But the words of the wise are quiet and few. --- > b...@yahoo.com wrote: > > Hi, [quoted text clipped - 31 lines] > page. This is really what you're trying to do with Audio -> MIDI > conversion. sounds like a good question for comp.dsp or for the music-dsp mailing list. if it's single notes, there is a possibility, but using the Fourier Transform to do it is likely not as helpful. try looking up "Pitch Detection Algorithms" in Google. also "Average Magnitude Difference Function" (AMDF) and "autocorrelation". r b-j > My first question is, how do I detext the attack points accurately? At > the moment, I take a fft of short length (256 for 44.1kHz audio), and > I add the values in all the bins and dividie this value by the largest > value in a single bin. This lets me know how noisy or chromatic a > single is at a point in time. I would think that a good discriminator can be the dwell of a constant frequency, ergo a "note". Such a dwell, or window of selectable width of from 100 to 1000 milliseconds, should make a proper "note" sensor. The FFT would be examined for an outstanding frequency subsequent the "noise" of an attack, whereupon said attack and dwelled tone are declared to be the new subject note up for analysis. > A guitar attack is noisy so this > information helps. But is there a better way? How about other > instrument playing single notes such as a flute? Add a filter acting as exponential time weighting to identify the note. One isolates ad identifies the attack as being broad band, while the sustained note sound following has much energy, most in its fundamental and a fair amount less in a few harmonics. The note emerges as identifiable as soon as a pure tone, can be identified in the FFT. The braod band noise occurs (attack), then in a few milliseconds the pure tone is the obvious survivor within the sound emitted. > In this case, I could > uses the fft to establish when a note is on or off by summing the > values in an fft window and comparing this value with some threshold. You would need a very fast Fourier transform. I hear that "wavelets" may be better suited, but that notion needs thorough investigation. > My second question is how might I calculate the loudness of this > attack, in sones perhaps (although since I am working with a wave file > were talking about relative loudness)? That's easy: Integrate all energy in an overall time interval that includes sustained not and the attack. It will take something of a long buffer to capture the entire attack and note information for post-note analysis. The end of that interval is marked by the next attack, not to be counted for the previous note. One can ten compute an attack-to-note ratio as the broad band energy divided by the total energy. Another common ratio is the attack to note ratio. Your choice. Angelo Campanella >> My first question is, how do I detext the attack points accurately? >> At the moment, I take a fft of short length (256 for 44.1kHz audio), [quoted text clipped - 49 lines] > > Angelo Campanella This approach will absorb a great deal of your time and teach you a lot. It will not, by itself, lead to a mechanical score-writing program. I have my doubts that such a thing is buildable, although in theory you could merely match any arbitrary selection of midi notes to a given sound signal, and eventally you might extract a midi score. FWIW a similar approach was used in modal analysis... and largely dropped. However, I do believe that a skilled composer with a full knowledge of the trends of his time could identify a melody, and parts, by listening to a multipart piece of music. Cheers Greg Locock |
Reply to this Message |
 Sign In
 Join
 My Latest Posts My Monitored Threads My Blog My Photo Gallery My Profile My Homepage Start New Thread Enable EMail Alerts Rate this Thread
|
©2009 Advenet LLC Privacy Policy - Terms of Use
This website includes both content owned or controlled by Advenet as well as content owned or controlled by third parties.