Piano World Home Page Forums Our Most Popular Forums Digital Pianos - Electronic Pianos - Synths & Keyboards The DPBSD Project!

I upped an MP3 for my Yamaha P120 if anyone is interested.

FILE NAME:
- dp_bsd_v1.3_yamaha_p120.mp3
PLAYED ON:
- Yamaha P120, Echo MIA soundcard, recorded with Adobe Audition.
PROS:
- 3 layers (from spectral phase view).
- Sympathetic resonance.
- Responds to partial pedaling.
- Huge dynamic range (~68dB, vel=1:127).
CONS:
- Looped, most loops sound wobbly at first, weak at the end, loop starts 1 to 3 seconds in.
- Weak sympathetic resonance.
- Obvious stretching, every three notes over low & mid, every 6 notes high.
- Layer switching is rather abrupt.
OTHER:
- Layer switch @ vel=88 and 102.

I have a P155 and I see you mentioned you would like that one on your list.

I'll try and find some time to have a go at the weekend.

Gerry that would be great! I'd really like to see how it compares to our old P120.

Instructions are in the readme file. Basically turn off the reverb, record the stereo signal as hot as you can without clipping, 44.1kHz sampling rate, save as 192kbps CBR MP3. I'm only interested in the main piano sound.

Cool, doesn't sound too tricky for an old IT geek like me!!

Oh, and the MIDI tempo should be set to 120.

dewster, I'm glad to see that you are finally putting your obvious technical knowledge to good use.

Initially, from the thread title, I thought you were planning to build a new software-based DP running BSD, but obviously 'BS' refers to something quite different!

Perhaps you will consider an open-source digital piano as your next project?

Anyway, I cannot promise anything, but I'll look into playing back your MIDI file on some modern KAWAI DPs (some support recording to USB memory as MP3 directly), if I get a chance.

Cheers,
James
x

Perhaps someone will make a UNIX bug detector and call it the BSD BSD project..


No way, DP customers are too picky! I'm actually much more interested in alternative controllers coupled to non-MIDI, modeled sound sources.


Sounds good, and I look forward to it!

Rendering MIDI to MP3/wave on a USB stick would be a super valuable feature in a good (solo recording quality) DP.

If it turns out you need a place to store files, you can upload them to Piano World.

The uploader is not just for graphics.

http://www.pianoworld.com/Uploads/fileuploader2.html

That's the "142 megabytes dedicated exclusively to high-quality piano waveforms" S90XS, I presume? I'd love to have a shot at that!

And would the RD700GX sample be from the RD700GX1 SuperNATURAL piano kit plug-in board? The one with seamless this and seamless that (everything is seamless)? I'd be super interested in that!

I guess it goes without saying that I'll squeal like a little girl the second I get my hands on a CFIII sample from the CP1/5/50!

Though not really on topic here, I just wanted to mention that the Motif XS can do this. One of the firmware upgrades added this ability. Really, you are "recording" to a sample and then writing that to an attached USB device (flash drive or HD). It is a very quick and easy process.

Tony

I did some tidying up, made some directories and sub-directories for the MP3 samples. Also created an upload directory for people to use.

Honing my craft while awaiting the CP1 sample.

Added my JV-1010 Session piano this morning:

FILE NAME:
- dp_bsd_v1.3_roland_jv1010_session.mp3
PLAYED ON:
- Roland JV-1010 Session piano, Echo MIA soundcard, recorded with Adobe Audition.
PROS:
- 2 layers (from spectral phase view).
- Large dynamic range (~60dB, vel=1:127).
CONS:
- Looped, actual samples are 3.5 to 0.5 seconds (lo to hi) with 1 to 0.5 seconds of crossfade.
- Looping obnoxiously static, decay tails sound organ-like.
- Stretching fairly obvious, stretch distances: 11,4,4,6,6,5,5,4,4,3,5,4,4,5,5,13.
- Layer switching fairly abrupt, switch @ vel=79.
- No sympathetic resonance.
- No response to partial pedaling.
OTHER:
- Anemic analog output w/ high noise floor.
- Fairly good for its time, now very dated.
- Date reviewed: 2010-02-05

One interesting thing about the output from this analysis is that it allows you to estimate the size of the physical memory required to hold the sample. Let's do this exercise for the Roland JV-1010 Session piano:

2 velocity layers * 16 actual samples / layer (number of stretch groups) * 2 sec / sample (rough average) * 44,100 samples / sec (sample rate) * 2 bytes / sample (assume 16 bit PCM) = 5,644,800 bytes.

5.7 MB.

Pretty freakin' tiny. You could put 175 of these sample sets on a 1 GB flash stick! Itself pretty tiny these days - the smallest I see for sale is 2 GB for ~$8. And at that rate, storage for the Session piano sample set would cost $4 / 175 = ~2 cents!

It's easy to understand why they played these games in the bad old days when ROM was expensive. Stretching in this case reduces the memory requirements by a factor of 88 / 16 = 5.5. Looping, assuming an average of 30 seconds is needed to fully capture a piano note, in this case reduces the memory requirements by a factor of 30 / 2 = 15. And 5.5 * 15 = 82.5 is a pretty whoppingly large factor to reduce the physical memory. Still, two fully unstretched and unlooped two-layer sample sets could fit on a 1 GB flash stick! A single four-layer sample set could fit on there too and would sound pretty fantastic in comparison.

I know the JV-1010 is really dated, but it's specs aren't that much different from the Yamaha P-120 (also out-of-production, granted, but less dated). These days I can see maybe playing these memory reduction games in a toy, but not in anything serious people pay serious money for.

With any luck we'll be leaving these bad-old-days behind us soon.

I'm redoing this post with PICTURES!

If people are interested in how I'm detecting layers, stretched notes, looping, etc. from the test file here are a few Adobe Audition screen grabs of the resulting MP3s:


The image above shows a spectral phase view of the stretch test (chromatic walk up the keyboard) of the Yamaha P120, where you can clearly see groups of three with identical phase - which means one sample is used to play all three consecutive notes within the group. As you might imagine, you can easily hear most of the transitions between groups.


Here you can see the spectral phase view of the velocity layer test (middle C from vel=1 thru 127, step=2) of the Ivory Bosendorfer, where you can clearly see 4 distinct groups of identical phase - so there are 4 velocity layers (at least for middle C).


Sometimes the velocity layers show up better using the spectral pan view, and in the above image you can see 8 distinct groups, hence 8 velocity layers in the Garritan Steinway sample.


In the above image you can see the spectral pan view of a note on the Yamaha P120. It starts out rich in harmonics, then about 3 seconds into the decay (@ the cursor) it turns into something that sounds wobbly and lamer - the dreaded LOOP! If you really think about it (and squint a lot ) you might be able to see the crossfade from sample to loop.


Finally, here is a clearer view of sample, crossfade, and looping on the Roland JV-1010, where looping is particularly egregious. The pure sample plays up to the cursor, the crossfade starts at the cursor and continues until the spectral view becomes largely constant approximately 1 second later - whereupon the crossfade ends and the pure, boring, organ-sounding looping is the only thing you hear, until either the key is lifted, or the loop hits the noise floor, or you go insane.

The sample itself is only 2 measly seconds long! And it's stretched over multiple notes! And I payed good money for this box, mainly for the Session piano patch! Granted, it was a while ago, but still...

Wow. You sure have been working at this project. As a "visual learner" I certainly find the graphics interesting and useful. I continue to follow this project, and am also particularly interested in the results for certain pianos.

Lawrence

I think I see a flaw in your system. You are not checking for how the sample sounds.

What if some one were to sample a $200 Craigslist upright that was not even in tune and has broken strings in a noisy environment using a Chinese knock off Shure SM58 mic. But he had tons of free time so he records all 88 notes with 128 velocity layers with each sample going for 5 full minutes and ends up wit a 10TB of samples.

I think your analyssi method would rate such a sample set as "the world's best".

Or for that matter what if I played you MIDI file on my very old FM synth based keyboard? I'm gussing you's score the old synthvery high, it never loops (or always loops depending on how you think about it) the velocity is handled with a continues filter, no steps and there is no note stretching. I think it would pass with 100%

But in both cases if yu tried to play music the over-sampled clunker and the Fm synth would both sound horrible.


I think we are seeing some of this effect already. I agree with your analysis of the Yamaha MP3 I sent in. You are saying (I think) that it is roughly as good as your P120. The sample is one dozen or so "free" pianos that come with Apples "Logic". And if you listen it sounds arguably much better than a P120.

Things like microphone placement and quality may matter more then the number of dynamic samples

So,... maybe you should include a midi file with a "best of" Beethoven exerts.. It would be fun to organize a blind listening test and see if the results match up to technical tests

First off, I must say I'm pretty impressed with this initiative by Dewster. I was wondering how he was able to tell apart all those things by ear, then the visual graphs blew me away, showing such a real clever, smart lad.

But I think ChrisA brought up a good point. I was wondering myself that the result of any modelled sound should be passing with flying colors according to the test parameters, because this test is sample-sound oriented, and modelled sounds don't have those same limitations, do they?

I understand why you'd want to remove the subjectivity part by coming up with such a technical test like this, but then all it will tell you is how detailed the sound was sampled and how much corners were (or were not) cut to get the final results. But you can't say much about the quality of the contents inside the waveforms, because that's the big subjective part that you want to cut out.

dewster,

I have a couple of Galaxy II pianos for you're listening pleasure but I still can't upload them directly to the new folder you created. From "my side" of your account I do not see any upload button like I have when I'm logged into my own account. I see the folder there but no way to upload it.

If you want the files they are in my folder where the others are. They are the Galaxy II Steinway and Vienna Grand Download Editions, hence the DE in the file names.

I'm not trying to pile on here, dewster, but with the vast knowledge of music here, classical or otherwise, it wouldn't take much to come up with a few suggestions for simple pieces of music or excerpts (for the sake of brevity) that you could incorporate into your test. (Have you heard this somewhere before? )

I think it would be very interesting to find out if the pianos that look so good "on paper" are actually the ones that we prefer listening to.

ChrisA, very good points! And no offense, but let me address them:


Well, people can generally go to Purgatory Creek for that. I'm not really that interested whether it sounds like a Steinway or not if the sample is only 2 layers / stretched like crazy / 2 second looped samples / 5 MB - by very definition that can't sound very good. If they put a ton of work into it and play a billion notes at once it might sound amazing for a 5 MB sample, but that's probably the best thing you'll be able say about it. The rest of the time you'll be wincing and wondering why it sounds like crap in a solo recording.

Look at it this way: say someone is telling you that they have the most amazing digital picture of something incredibly detailed, but the picture size is only 50 KB. Without even looking at it you can tell right off that you probably won't be very impressed. If it's a 4 MB file you'll probably stop thinking about technical issues and take a look. Same thing.


It would sound like an incredibly detailed sample of a POS, something you can't do with a 5MB sample no matter how hard you try. We're talking detail here. Once the ability to reproduce detail hurdle has been passed, then I start caring about tone. Before that, I don't really care, by very definition it's not a serious instrument, it's a toy.


This is more of a go / no go test for crap compression issues with the sample set, not a BUY IT NOW! test. People should know how much sample they're actually getting IMO as up till now at least it has been a huge determinant in terms of sound quality.

And a single layer in a sampled DP is usually a bad thing, and I would put it in the CON column.


I went back and looked at your sample again yesterday (I'm slowly figuring this out), let's compare them again:

------------------------------------------------------
- Unknown small sample set from unknown Yamaha piano -
------------------------------------------------------
FILE & SETUP:
- dp_bsd_v1.3_yam_studio.mp3
- Unknown setup.
PROS:
- Looped, but fairly well done.
- OK sympathetic resonance.
- 2 layers with some kind of filter? (from spectral phase view).
CONS:
- Rather rapid note decay.
- Poor dynamic range (~30dB, vel=1:127).
- Looping of higher notes fairly obvious.
- See evidence of at least 2 note stretching.
- Abrupt layer switching.
- Does not respond to partial pedaling.
OTHER:
- Volume in MP3 file too low! (please try for -1dB peak)
- Date reviewed: 2010-02-04

----------------
- Yamaha P-120 -
----------------
FILE & SETUP:
- dp_bsd_v1.3_yamaha_p120.mp3
- Echo MIA soundcard, Sonar 6, Adobe Audition.
PROS:
- 3 layers (from spectral phase view).
- Sympathetic resonance.
- Responds to partial pedaling.
- Huge dynamic range (~68dB, vel=1:127).
CONS:
- Looped, most loops sound wobbly at first, weak at the end.
- Weak sympathetic resonance.
- Obvious stretching.
- Layer switching is rather abrupt.
OTHER:
- Samples are 3.5 to 1 seconds (lo to hi) with 1 to 0.5 seconds of crossfade.
- Stretch distances: 8,3(x25),5 = 31 groups.
- Layer switch @ vel=88 and 102.
- Date reviewed: 2010-02-04


So the free Apple piano is probably a 2 layer (w/ filter?) with rapid note decay, poor dynamic range, and doesn't respond to partial pedaling. This exercise isn't necessarily to pit one DP against another, but more to expose what is going on in each. That said, the P120 definitely has better specs, and sounds better to me (IMO). And in this case I don't believe that's a coincidence.

Thanks setchman! I snagged the files and will analyze them and upload them to the archive tomorrow. I can log out and still upload to my folder, so I'm not sure what's going on. Others are having the same problem, so it's not just you. I guess as long as people can post them somewhere where I can get at them that's the main thing.


I hear you, and do agree. But you can post samples to Purgatory Creek for sound issues, I guess I'm more interested in what's holding the vast majority of DPs back from being serious instruments at this point than what they sound like. Hopefully soon that will be just the opposite.

I hope I addressed this concern somewhat above - the more layers the merrier in a sampled DP; no layers are the norm for a modeled DP - and both should be judged from that perspective. I did test the TruePianos demo, a supposed sample / model hybrid. It had weird phasey stereo issues in the decay and obvious stretching. Any approach can sound either good or bad, particularly individual notes.


This isn't an attempt to remove subjectivity of sound in any way. It's more of an exercise in cutting though the hype - a test to see if it is even possible for a given DP to sound good. Up close most of them look and sound fairly ugly.

Most of the PC-based samplers & modelers I've tested so far I would be very happy to have in a DP.