In the following thread, I tested ATRAC Type-R's performance on a castanet sample. At the request of Guruboolez, I've uploaded the Type-R encoding and the original, both edited to be the same length (and normalized).

((EDIT: Please scroll down to KikeG's reply for an edited download of the clip))

The ATRAC-R sample phase is inverted compared to the original.

Edit: also, the left channel of the ATRAC-R sample is 0.7 dB louder than the original.

Edit: removed "warning". from message, see below.

If that was the only difference between two samples, would it be ABX'able in itself?

Really, it shouldn't. I shouldn't had written "warning", sorry. But the 0.7 dB channel difference maybe could be ABX'able by itself. (*)

(*) Edit: I don't know if replaigaining fixes channel imbalance.

I hope not. Surely, if a change is made by replaygain it's done to both channels.

Hmmm, that's interesting. I wonder what happened when I capured the sample in Audacity? Unfortunately, I don't have a digital input on my laptop soundcard, so the capture had to be via the line-out of the recorder (set on 'line out" mode to bypass the internal amp).

In any case, I've checked the ATRAC sample and I can't perceive a difference in volume or phase, but that's definitely a (possible) confound variable in the tests. I also guess it depends how you attempt to ABX the sample...I wasn't listening for a "difference," per say, but a specific type of artifact between 2 and 3 seconds. That's probably also why I didn't notice the smearing that Guruboolez heard on the guitar part of the passage.

If the sample is not satisfactory from a research standpoint, would somebody else with better equipment be willing to encode the castanets sample digitally and upload it?

Here is the ATRAC-R sample time and level-aligned, both channels. There's not much difference from my point of view.

Thank you, Kike G.

For the future, what might I do differently to avoid this from happening again? Or are the problems possibly a function of the line-out of the recorder and line-in of my soundcard?

(Edit: I removed the mis-aligned version from the top of the thread and directed readers to the corrected clip)

How about ABX'ing ATRAC3plus? Bitrate is a bit lower though at 256kpbs instead of 292
Right now I don't have any sophisticated wav editor so I can only record and trim...

I can't personally test ATRAC3plus because a) I am on OS X and Sonicstage crashes when one tries to encode under VPC 6.1, and b) I don't own a "Hi-MD" recorder and wouldn't buy one anyway . So, if you'd like me to ABX some samples, you'll have to encode them in SonicStage 2.0 and upload them to the board. It would make for an interesting test, though.

Like I said, I can upload a non-corrected sample
Anyone can "fix" it for me?

It's possibly your sound card or the MD inputs, or a combination of both.



I'm posting here your original castanets sample, because it's time and level-aligned with the ATRAC-R sample, and the PCABX sample is not.

I can.

Ok, but I guess my question is a) using Audacity, how might I detect these errors, and b) how might I correct them. I wanted to encode a bunch of killer samples, but I need to figure out how to get accurate samples from my equipment.

Thank you very much for all your help with the samples, posting new ones, etc.

The castanets.wav sample encoded with SonicStage 2.0 on ATRAC3+ at 256 kbps.

Note that this file was generated by dlb for his unlucky listening test:
http://www.hydrogenaudio.org/forums/index....c=20485&hl=test

ATRAC3+ at 256 kbps is replacing ATRAC (type-R) at 292 kbps on modern MD device.

I'll put the results of my ABX in a seperate thread. ATRAC3plus was exceedingly easy to ABX, 20/20. Sony, are you listening?

a little off topic. what about the atrac dsp type-s?

ATRAC Type-S is just a combined, one-chip DSP for the ATRAC Type-R codec used for SP mode and the ATRAC3 codec used for MDLP modes, so there's nothing new to test. Supposedly, Type-S increases the SNR of LP2 and LP4 encodings by 6dB upon playback, but doesn't offer any improvement for SP encodings.

I use Cool Edit Pro, now called Adobe Audition, but I'm afraid I haven't ever used Audacity. CEP has a feature called "synchronize cursor across windows", which means that when you have several waves loaded, they all are represented on same time-basis, that is, using same time-axis zoom, and same zoomed area. This helps quite a bit on the process, but is not essential.

First thing you need to do is verify that samples are same phase. That can be done by looking at both waves on time domain, with low or no zoom, on some parts containing non-symmetrical shapes, and comparing similarity of those shapes "by eye". Of course, they must be same parts on both waves. Non symmetrical means that the part of the wave that is over zero is distinctively different from the part that is below zero. Look and compare at different parts of both waves, in order to verify that those shapes are similar, or are inverted vertically. If phase is inverted in one, just invert the whole sample so that phase is the same in both waves, and verify that shapes are now similar.

Now, once phase is the same in both samples, you must time-align both samples with precision. I believe ABC/HR Java tool can calculate time offset difference on two samples, so this might help you. However, I don't use it. I time-align samples also by hand, by zooming some characteristic part of one of the samples (some peak that is easily identifiable, for example), and then, locating that part in the other sample. Then, I calculate the difference in samples using the cursors (thanks to the fact that both samples are time-synchronized in CEP). Then I see which sample is delayed with respect to the other, and and delete the calculated nº of samples in it, so that the delay is reduced to zero. Another way of achieving this is deleting some samples "by eye" on the delayed sample, several times, until it is no more delayed with respect to the other. Ideally, this process should be done at last with a precision of just 1 sample, zooming the waves as needed, but if the precision is of a few samples if can be ok too.

Now, the level-align issue. For this, after some experimentation, I have found that the more reliable and easiest method is what follows. First, you need to create, with the time-aligned samples, two new wave files, one with both left channels of the two samples to level-align, and another with both right channels. I do this by simply using copy & paste on a new blank wave file (of the same freq. and nº of bits than the samples at issue). So, I create first the left-channels new wave file. Then, I perform a FFT (spectral view or "Frequency Analysis" in CEP) of this new sample. Then, I can see a frequency graph of both left and right channels of this new sample at same time, being those left and right channels the left channels of the two samples to align. In order to have a good picture of levels vs. frequency in both channels, I use logarithmic view ("linear view" not checked), a FFT size of 1024 points, a Blackmann, Blackmann-Harris or Hanning window (this doesn't matter much), and I average ("scan" in CEP) a portion of the wave with some loud content, in order to have an averaged, smoothed frequency graph. In case of castanets sample, I scanned the 4 castanets part of the sample. I use a dB range of around 40 dB or less, for a good visualization of small level differences. Then you should see that left and right channel graphs have nearly same shape, but being one of them displaced a few dBs of fractions of dB over the other. If they totally overlap, they are same level. However, for having best resolution of level differences I just take readings from the cursors at a several parts of the spectrum that are high amplitude, between say 100 and 5000 KHz, and just calculate the difference. Difference on those parts should be the same, or very similar. That difference is the level difference between both left channels of the samples, you must annotate it. If level difference is below or around 0.1 or even 0.2 dB, then it's ok, no corrections are needed. Then, repeat same process for right channels, and annotate difference. Of course, you must take note of which sample is louder, and such.

Here I post a screenshout of this last step. There's a 0.4 dB difference between both left channels of the samples at issue (in the graph, between left and righ channels). The reading is at 366 Hz.


Now, to correct for those differences, you must open the sample to correct, and amplify/attenuate level of its both channels (separately if needed) so that they compensate for the differences found. Best way to do this is to convert the sample to 32 bit, do gain (level) correction, and then convert back to 16 bith with dither. I use flat (no noiseshaping) dither, triangular pdf, and 0.9 or 1 bit of amplitude.

Of course, it's a good idea to save the wavs just after every step of correction is performed, so that you can go back if you need to.

So, the whole process may take somewhat for every sample processed, but once you have some practice, it takes less.


I just wanted to help, thanks.

Wow....that's a lot of info. Thanks! Now I have to figure out if this stuff can be done in something like Audacity. The problem is that I'm on Mac OS X (and am very poor, being a student), so I am pretty much limited to it as opposed to payware. If I upload any more samples, I'll be sure to follow your advice and hope that you'll take a look at them to make sure they are ok. I've already found a few samples that I may want to try out, though the outcome is pretty much assured already

I'm getting the impression from the many negative reviews I read of NetMD and the newer Hi-MD (in addition to the fact that a search on many MD forums shows many iPod converts) that MiniDisc is in its twilight as a format, so any further testing is just to quell people's curiosity.