Hello,

I want to transfer a huge LP collection to CD-R for occasional listening, ...in my office, car etc.

I bought external M-Audio Audiophile USB ADC/DAC

It can record up to 24bit/96kHz

My TT rig is:
Thorens TD-126 MkI turntable,
SME 3009 S2imp arm,
Shure V15VxMR cartridge,
Graham Slee Gram Amp 2 SE phono preamp,

My SME was rewired with Van den Hul MCS - 150 M cable.
Cable from tone arm to preamp is Van den Hul D - 502 HYBRID.

Shure V15VxMR has this microridge stylus, which is great for archiving old LPs - much less distortion from worn out grooves.

The downside is that the treble is rolled off and the sound is too warm for my taste.

(yes, I played a lot with cables and capacitance, but this just seem to be the character of this cartridge)

Using test LP sweep tones, I created a custom graphic EQ preset for Adobe Audition 1.5 (CoolEdit originally before Adobe bought it).

Now I get ruler flat frequency response from my cartridge/cable/preamp combination and the sound is much crisper now :-)

The big question is:

do I record at 24bit/96kHz, apply the graphic EQ and then downsample to 16bit/44.1kHz?

At first I thought that this is a good idea.

But then I started reading numerous web pages dedicated to downsampling...

My brain got swollen from all this dithering, triangular noise shaping, custom curves, bit mapping, etc.

I came to conclusion that I might get some distortion, additional noise, lowered dynamic, digital artefacts, etc. when I downsample.

On the other hand, 24bit has much more information to work with...

I only plan to use the graphic EQ with the adjustments no bigger than +3.8 dB.

Nothing drastic.

Do I get better sound by applying moderate EQ in 16bit/44.1kHz,
or doing it in 24bit/96kHz (or better a 2x multiple of 44.1kHz) and then downsampling to 16bit/44.1kHz

If downsampling is better, what settings would you recommend for Audition/CoolEdit?

Thanks a lot,

Aleksandar

I'd use 24/96 and then when you're finished doing adjustments resample using a plugin or external resampling program based on SSRC. (which is widely known to have the best quality resampling techniques, definitely inaudible anyway.)

The answer, to satisfy yourself, is to create some samples and do a little ABX testing with WinABX or PCABX. What you will find is that it does not make any difference to the end result. Especially it does not make any difference to work at any higher sampling frequency if you are going to make CDs.

If you are going to do much processing besides the EQ, such as clean out the clicks and reduce the broadband noise, there is some advantage to working at 32 bit, but with all the noise inherent in the LP it will be hard to tell the difference in the end. There will always be rounding errors if you work at 16 bit, but they won't amount to anything you will ever hear unless you preform quite a few transforms.

Using test tones, it is easy to demonstrate that your card produces a significant aliasing image. If you record at 44.1kHz, a input sweep tone that goes to 30 kHz or higher will make an aliasing image that is almost as strong as the input signal for the first few kHz below the 44.1kHz Nyquist limit. By recording at 88.2 or 96kHz you mostly eliminate that image from the 44.1kHz space. Anything that might exist at higher frequencies is well removed when you downsample.

Very little music will have much signal strength above 22,050Hz however, so its aliasing image will be very weak to non-existent, probably totally insignificant. So, while what I described above is real and easily demonstrated, it is of little significance for most LP transfers -- few have strong enough signal levels above 22kHz for you to ever tell any difference. There is no other reason what-so-ever to record above 44.1kHz if your goal is CD. Anything at all that might be gained will be totally lost when you downsample to 44.1 for your CDs.

Audition does an exceptional job of resampling, so if you do record at a higher frequency or greater bit depth, you do not need any other program for resampling.

Also, unless this USB card is USB 2, and actually makes use of USB 2, it cannot record at those higher frequencies or bit depths in Audition. You need ASIO drivers to accompish that under USB and Audition does not support ASIO.

>24bit/96kHz, apply the graphic EQ and then downsample to 16bit/44.1kHz

You are proabbly best recording 24 bit 44.1KHz, keep the frequency the same, apply the EQ then reduce the bit depth, I don't think there is anything to gain by going 96KHz if the end result is 44.1KHz.

That is a tricky one. If you set fs = 44k1, you should provide good filtering before digitisation. Vinyl does have a lot of harmonics above 20 kHz, mainly distortion components. One should avoid aliasing.

So, record in 24 bits/88k2. Do your EQ and normalize to -2 dBFS (to avoid clipping between two samples due to cheap interpolation filters). And convert this to 16 bits/44k1 with proper dither: triangular pdf noise with 2 LSB bit peak-to-peak. At least that is how I would do it.

Regards,
Jacco

All ADC used today are delta-sigma (or variants), with huge oversampling first (Nxfs, where N >= 256), then digital filtering to cut freq above fs/2, then picking every N sample. This means that the (analog) filtering needed is to cut frequencies above Nxfs/2 (which for N=256 and fs=44k1 is ~ 5.6MHz). A simple two pole active filter at fc=40k (meaning almost zero attenuation and phase shifts in passband 0-30 kHz) will have >84 dB attenuation at 5.6MHz.

I'd say go right ahead sampling at 44k1.

-r

Most of the cards I've been able to test specify 64X oversampling on A to D. One might think that would completely eliminate aliasing images, but as I wrote on the 18th, it is very easy to demonstrate this isn't so. Quite strong images are produced from test tones when recording at 44.1kHz. Of course real world music will have much lower levels, if anything, above 22kHz, so the aliasing images will be correspondingly weaker.

However, I can handily record the 1kHz to 30kHz sweep signals on the Cardas Frequency Sweep and Burn-In Record, along with multiple harmonics of same that go as high as 96kHz will sample (obviously, 48kHz). This is direct evidence that fairly strong higher frequencies can exist on vinyl. My tests show such signals, when they exist, will produce aliasing images that will be almost as strong in the 18kHz to 22kHz region as the signals themselves are in the few kHz above the Nyquist limit.

This is true. Many ADCs have a filter that is just -6dB at fs/2, so they let in quite a bit of aliasing around fs/2. In order to avoid this, a good ADC with a better filter, or good software resampling, are a must.

However, as has been said, in practice the audibility of those aliases will be very low, and most probably non existant at all, mostly due to their high frequency and low levels.

Could you upload examples of this?

I was going to post some screenshots that show the aliasing images but this forum doesn't appear to let me attach anything. If I have to have a web page when I can put such things and just point people to it, I'm afraid we are out of luck. If there is some other way, perhaps someone can point me to the instructions.

Try http://www.putfile.com/

I'm guessing this is what you want to see. I did not keep very much of the test material I generated when I was playing around with this, a couple years ago, but I probably have this because it is a reasonably representative. I have put the images on putfile, as suggested, but I'm not sure how things will turn out. When I go to the page with the images, all I can see or download are very low resolution thumbnails. They are almost useless for out purposes here, but maybe this posting will be somehow different.

Both are screen shots of Spectral View in CoolEdit. Input was a sweep tone ranging from 8kHz to 48kHz over ten seconds. The pictures are shorter segments that highlight the more intense parts of the alias image. The range was increased from the default 120dB to 160dB so that the fainter trace would also show up. That fainter image can be seen to go back down to 8kH when viewing the full duration.

The alias image is that part to the right of the cursor. It looks not unlike a reflection off the ceiling. It can be seen that either my memory was a bit faulty or this isn't the most extreme case; the higher intensity part of the image is fading below 20kHz rather than persisting strongly to 18kHz.

When recording is done at 88.2kz or 96kHz there is no alias image. There probably would be an image if the input frequencies exceeded 1/2 the recording sampling frequency, but the strong parts of that would be up near 44kHz or 48kHz. When this input signal is recorded at the higher sampling rates, then properly resampled to 44.1kHz by CoolEdit, the result is completely clean.

This second image clearly demonstrates the problem recording at 44.1kHz with Creative soundcards that resample everything to 48kHz internally. There is a lot of extra junk from the internal resampling. A 48kHz recording looks much the same as those from other cards. (The bright vertical lines are clicks due to some problem I did not try to solve. I don't believe they are relevant to the item of main interest.)

I ‘ve had only a small sample of existing soundcards to test. Possibly the best of those was the Echo Mia, which is widely, if not universally, consider to be adequate for professional recording. All cards tested gave more or less identically results. I would be quite interested to see tests results from more expensive cards, such as Lynx or Apogee.

This is track 2a from the Cardas Frequency Sweep LP


Nope, still seems mostly useless. I thought I read the putfile instructions but maybe I'm missing something.

I see that the thumbnails now provide a larger version of the screen shots when clicked upon. What I see still is not as clear as the images I loaded onto that Putfile site but they are better than the thumbnails here.

What devices did you use for the the playback of those sweeps and for recording?

I created the sweep tones in CoolEdit in a 88.2kHz sample space (we are not talking about the Cardas LP of course). I played them in looping mode (just to allow myself plenty of time to record all I wanted) to the analogue out of an Audiophile 2496 which was cabled to the analogue in of another Audiophile 2496 on another computer.

On the second computer I recorded in CoolEdit. If I recorded at 88.2kHz or 96kHz, the recording essentially matched the created sweep tones. When I recorded at 44.1kHz (or 48kHz) I got the aliasing results displayed. The screen shots are of selected portions of the recordings displayed in CoolEdit's Spectral View.

I ran the experiment in both directions to test both Audiophiles.

The SoundBlaster recording was the same except that the receiving audio card was a SoundBlaster.

I was also able to test an Echo Mia and an M-Audio Duo. Both produced results matching the Audiophile.

The Cardas Sweep track was recorded using a Shure V15VxMR, which is surely the limiting device in that reproduction chain.

I've been recording LPs for transfer to CD-R for some while at 88.2kHz. I resample to 44.1kHz before further processing. Occasionally an LP will have almost nothing above about 5kHz but others show strong color (Spectral View) all the way to the 44.1kHz Nyquist limit of the 88.2kHz recording (Well, it is at least fairly bright to 40kHz or so and still visible at the top.)

?

The V15VMRx is spec'ed as follows, according to Shure
http://www.shure.com/phono/v15vxmr.html

"Essentially flat from 10 to 25,000 Hz"


Now, it's possible their definition of 'essentially flat' is very unusual, or that you got a defective cart, or it's not set up correctly. How did yours measure?






well, as you probably know, 'downsampling' refers to the sample rate, not the bit depth. If you're worried about downsampling, trye using a multiple of 44.1 -- e.g., 88.2.



+3.8 can sound quite different depending on where it falls in the frequency range...it's at least possible too that some signals could be driven into clipping, if you don't leave yourself enough headroom.





Theoretically you *could* introduce artifacts by doing digital processing at 16 bits but it would have to be extensive rounds of digital processing. But with LPs any such 'lost' dynamic range is pretty moot. If you want to do a higher sample rate capture, I'd recommend 88.2 (2x multiple) as above.

When you do a frequency plot of such captures, what levels are you seeing above 20 kHz?

The answer is not exactly straightforward, but the levels are very low regardless. Probably low enough that any alias image is almost certainly no more than a mathematical fact rather than an audible fact.

Any numbers associated with a Frequency Analysis graph are dependent upon how the measurements are made and also very dependent upon which recording, and just where in the recording one chooses to make a measurement. If I choose some mid 60's surfing music, with lots of electric guitar, Spectral View shows signal right up to the 44.1kHz cutoff (when recorded at 88.2kHz). I could pull out some excellent piano solo recordings made a decade and a half later that show very little above 8kHz.

Looking at a Frequency Analysis graph of some of that surfing music, I find a more or less smoothly declining curve from about 12kHz onward. I normally use FFT Size 16,384 for Frequency Analysis. This produces more point to point variation than a lower FFT size and lower numbers for any particular point. A higher FFT Size of course produces still more variation and corresponding lower numbers. The individual numbers are not indicative of the signal level one would measure at the output of a DAC. That signal level is a summed measure of all the FFT windows across the entire frequency range.

The two lines below are from using FFT size 16, 384. The first of these is a short selection in the vicinity of one of the many frequency peaks. The second selection is about two seconds duration surrounding the first selection.
short selection -- 12kHz: -61dB 44kHz: -118dB
longer selection -- 12kHz: -66dB 44kHz: -116dB

These next two measurements are of the very same selections but using FFT size 128 for the Frequency Analysis graph.
short selection -- 12kHz: -34dB 44kHz: -82dB
longer selection -- 12kHz: -43dB 44kHz: -93dB

It would be a very tedious measurement and calculation process to determine the contribution to the signal level of the components above 20kHz, at least by means I have available. However, if I take the 2 second selection and apply a 20kHz cutoff FFT high pass filter, the RMS Average value over the two seconds drops from -20dB to -60dB. I can't hear a thing if I play it, of course, but If I adjust the sample rate downward, say to 11025Hz or 8000 Hz, that remainder is readily audible at normal listening levels.

As some limited reference to people not familiar with these tools: I just did a noise floor reading on this computer. The Average RMS value is -98dB (CoolEdit Statistics). A Frequency Analysis graph (FFT Size 16, 384) shows a fairly level line across the frequency spectrum. Any particular place I put the cursor reads about -132dB. That -98dB is the sum of all the (approximately) -132dB samples.

You can actually select and scan the entire file, in Audition. Default 'resolution' (FFT size) for scanning of levels, is a sample every 20 Hz or so, if I recall correctly. I presume the software averages the levels at each sample frequency, across the whole scanned area. You can copy the actual values from 20 Hz to 22 kHz to a clipboard in Audition and paste them into another file for viewing. I do this to compare various different masterings of the same track, for example.

You can certainly spend the time doing a scan of an entire track, or an entire album, but generally you will only get a slightly more 'rounded' view of what you see from a couple seconds or so. There won't usually be much change going from a 2 seconds to 400 seconds.

This isn't true if there are major variations over time within the track -- now a full orchestra, now a bass drum solo, now ..., where you could get different graphs depending upon just what you select. That was somewhat the case when I selected only a small section where Spectral View showed that levels reached the top over the entire duration of my selection. This made the very high frequency averages higher than when looking at an extended duration where sometimes the highest frequencies present might be only 6kHz. In the case where there are significant variations over time, I'm not clear what could be gained from a full track scan.

I know the latest version of Audition has some additional facilities under Frequency Analysis. Perhaps one day I can upgrade from CoolEdit, but I have to get along with what I have for now.