Piano World Home Page Forums Our Most Popular Forums Piano Tuner-Technicians Forum Note detection in piano tuning software

Hi,
I recently tried using UltraTune (the freebie) to do a "patch up" tune on my Yamaha baby grand (I am not a qualified piano tuner).
I found that it was not able to reliably track notes outside G3..B5 (on my piano, in my studio, with my microphone, etc...)
Being a software designer with decent signal processing knowledge, I figured I could do better, and so far I am getting some very encouraging results.
Except in the top octave...
My question is: for users of Veritune, how good is it at detecting pitch in the top octave?

It usually works fine. Except for C8, A0, A#0 and B0 in some pianos.

I usually have no trouble, all the way to C8 - the low bass with some pianos is another story.

There are tricks...

Lock the display on the desired note
Repeat the note - more than once a second...
Move the unit - sometimes an inch or two makes the difference.


As to the programming stuff? Sorry, no help here!

Ron Koval

You have to crank recording levels up and move the Mic to treble or bass. But yes, it has detection issues.

One of the difficulties in making the note detection too sensitive is that those little sounds that a hammer makes when it touched a tuning pin makes the whole plate resonate. Such sounds can appear very much like high treble notes. If the detection algorithm is too sensitive, it will switch notes sometimes when these noises are made during tuning. Since it is almost impossible to silently move your hammer from one pin to the next, this can be quite a nuisance. That is why note switching algorithms have to place just the right amount of qualifications on signal characteristics, and it is these qualifications that prevent perfect recognition of every new note.

I don't have problem with verituner. For tunelab, need manual note switching for most notes.

Detection of low note need special algorithm. For those low notes may have weak fundamental even missing. For high treble, it decayed in second. You have to design when to sampe it. For aural tuning, just listen to attack to decid if the frequency is correct.

Program seems simple to most programmer. Piano acoustic
is
not easy to understand.

RCT has similar issues. If I set the parameters to where it auto searches/moves in just the upwards direction, 1/2 the false hits disappear.

I've put version Pitchfork v1.0 on my website: www.cloneensemble.com.

This won't be serious competition for the big boys - either full price software packages, dedicated devices or professional aural tuners. But it might be a decent alternative to some of the other free offerings.

The next feature on the list would probably be a lock-in note selector, as pitch detection is the #1 problem in the extreme octaves.

I've had good luck with this mic, I436 in detecting higher notes on the piano http://www.mic-w.com/product.php?id=3

A couple of vids showing how it works with my Ipad 2 Verituner.

http://youtu.be/nLnO6YeCaDM

http://youtu.be/cGXCszP-c6w

Trevor. I can't run your program because I do not have a PC (I am Apple-bound at present), but I can see the main window image from your website.

I am interested to ask you a few questions because I also do a little programming but not to your level. I have dabbled with pitch detection algorithms whenever I get some spare time, but nothing good enough so far that I would let to the public.

Two basic approaches that I have successfully used to detect pitch have been:

1) A continuous strobe that samples the incoming audio stream at periods defined by the pitch requested and the audio parameters. If the audio input has the same frequency and period as the defined period length then the audio is sampled at the same point in the waveform and will be stationary. If the audio waveform varies slightly then the sample will look at progressively different points along the waveform. Various display method can be used to represent the shift such as moving squares etc. The accuracy of this approach is very high and limited by the duration of the note during it's decay and can be easily down to 1/10 cent levels.

2) Recording audio samples for a defined duration ( e.g. 200ms) and processing the waveform with some sort of transform algorithm to get a frequency spectrum. The spectrum is then inspected for peaks that represent frequencies sounded. The transformation could be at 1/10 Hz intervals, but as longer the sampling time is used, and if a wide spectrum is analysed, then there can be too long a time period between sample taken up by processing and that interferes with the flow of tuning. Ways I have used to make the process quicker is to start the spectrum with a wide range of frequencies and analyse at 1 Hz intervals to get approximate frequencies, and then re-do at 1/10 Hz across a then selected narrow band to get greater accuracy. Another problem is what to do with the beginning and ending samples that have incomplete audio information.

With your screen image I see that you have a railsback curve. Does that mean that you are detecting the full range of piano notes at their fundamental frequencies, or do you do as others do and decide to listen to specific partials for the lower notes. That would lead to the railsback curve becoming a "tuning curve" with steps in the curve in the lower have or so as the preferred partial takes over.

It has been fascinating for me to dabble with this sort of stuff. I wonder if you could elaborate on how you go around detecting pitch?

Edit: I remember now that piano string inharmonicity is the major problem because the audio period does not repeat itself the same each time. Detecting nice sine waves is easy, but complex piano waveforms are difficult. I have experimented with pre-filtering the audio waveform as a means to accentuate particular partials for analysis, but my filter algorithms are not quite good enough yet, and this concept assumes an approximate prior frequency estimation.

Hi Chris,

sorry for the delay replying. The approach I have is FFT-based, but I don't do anything with the resulting spectrum - that is far too coarse.

Instead, I "push further" with the FFT spectrum and perform an auto-correlation - that is, effectively slide one copy of the waveform against another, and compare how closely they "match up". The "displacement" (amount of slide) of the closest match corresponds to the wavelength, and thus frequency, and the precision is at least as good as the sampling interval (with tricks to make it even better). With care, this gives a more precise result.

The pitfalls are that some piano notes have a very strong octave harmonic, which means they can lock on to that (giving a match an octave high). Conversely, some notes have a "quirk" that affects every second cycle - you don't hear it as a sub-octave, but the autocorrelation detects it, and gives you a note that is an octave too low. Both of these can be alleviated by (heavily) processing the waveform before the correlation - filtering, centre-clipping, sigmoid-based distorting. But the amounts of all of that are very tricky to arrive at, and (frustratingly) depend on what note you are dealing with in the first place!

Finally, the extremes: very high notes are of such short duration that there is very little "space" to slide the copy. (And their "thud" is rich in confusing low frequencies). And in very low notes the (horrendously atonal) harmonics are far stronger than the fundamental, especially on smaller pianos, giving a pretty much random autocorrelation.

I've spent quite a lot of time trying to fine-tune my algorithms, based on single-string recordings of my Yamaha baby grand. It's not perfect, and I have no idea how it will translate to other pianos, but using those techniques I did made some encouraging progress. Knowing what I know (and what I don't) I'd leurve to know some of the tricks the big boys use (eg Veritune) but I don't think they are going to tell me