Wednesday 1 July 2015

ANALYSIS: DSD Conversion Impulse & Spectral Display... Another few images for completeness!

A cool example of the "Spectral Frequency" display - from Adobe.

Okay guys, I actually wrote the text that follows 2 months ago before the Digital Filters Test. As happens sometimes, a post can get lost in the "draft" bin and later found. It refers back to the DSD-to-PCM analysis series from April:
ANALYSIS: DSD-to-PCM Conversion 2015 - Windows & Mac OS X
ANALYSIS: DSD-to-PCM 2015 - foobar SACD Plug-In, AuI ConverteR, noise & impulse response...

Consider this as part 3 of the 'trilogy' for this week as I continue to work on the results of the recently-closed Digital Filters Test.


I was reminded recently by Wombat in his post here that we can have a look at the spectral frequency display as well when assessing sonic data in the audio editor. Also, Mnyb talked about the low-pass filtering in SACD players and what "standards" were used. Well, I don't know about the formal standards, but the first Sony SCD-1 SACD player back in 1999 had a defeatable analogue filter placed at 50kHz and according to the DSD Wiki, all SACD players were supposed to include this "optional" filter. I know a few people on audiophile forums suggest they subjectively preferred the filter be turned off.

Anyhow, here's what a simple 10kHz PCM sine wave originating as 24/176.4 would look like after being converted to DSD64 and DSD128 (again with Saracon) and then decoded back either totally unfiltered with foobar2000 SACD plug-in and also if we applied a 50kHz low-pass filter with iZotope RX 4. Overlaid are the 3 typical views - waveform itself, FFT frequency analysis, and spectral frequency display using Adobe Audition:

In each of the spectral frequency plots, we see the bright yellow 10kHz signal as well as any noise within the spectrum up to 88kHz for the 176kHz PCM, and 176kHz for the DSD conversions done at 352kHz. Of note, the DSD128 conversion with a 50kHz filter looks pretty good. DSD64 with 50kHz (simulation of the default Sony SCD-1  / SACD player filter) isn't bad but obviously we see the noise >20kHz breaking through a bit and if you click on the image and look at the sine wave closely, it's not hard to see a few "imperfections" from going through the DSD64 encoding stage.

(For practical reasons as I really don't think we need to concern ourselves with noise >96kHz, each case was analyzed at 192kHz instead of the maximum 352kHz output from the DSD decoders.)

The spectral frequency display also beautifully demonstrates to us how the impulse response differs for each of the DSD-to-PCM decoder programs. With a standard PCM impulse run through a typical sharp linear antialiasing filter, we get the top image. The spectral display demonstrates nicely that the "ringing" is at the Nyquist frequency; 48kHz.

The second image is the nastiness of DSD64 noise without any filtering at all (from the foobar SACD plug-in)...

Below that is Saracon DSD64 decoding. As discussed before, it uses a ~30kHz filter. There is a little bit of noise that seeps through but really quite minimal and overall a nice compromise. Pre-ringing is minimal in comparison to the top image.

JRiver 20 with the "Safe" 24kHz low-pass filter utilizes an intermediate phase setting and we see a slight asymmetry to the spectral plot showing mildly accentuated post-ringing.

DSD Master shows us the post-ringing skew from a minimum phase filter. I don't think I mentioned this before, but it looks like DSD Master implements a relatively steep 50kHz low-pass filter to remove all the other high frequency noise.

Finally, the AuI ConverteR can be seen below with the steep linear 20kHz filter. Spectral display demonstrates the symmetrical "ringing" phenomenon at 20kHz of relatively long duration due to the steepness of the filter.

I think I'm finally done looking at these DSD decoders! :-)


To end off this week, I've noticed that there are still a number of very suspect "high resolution" recordings out there released in 2015. For example, recently there has been a release of Andrea Bocelli's discography at the usual vendors. For example, here are some pictures from the album Amore (2005) of the 24/96 release:

This is the song "Besame Mucho", track 2 from Amore. The top image is the usual waveform display which shows quite a significant amount of compression - not a surprise for pop albums, including these classical/vocal/pop crossover genres. In the lower spectral frequency display, we see something unusual going on... There's a "band" of low energy centered around 24kHz. If we have a look at the FFT where the yellow cursor (left side of the image) is, we see this:

There's a rather large symmetrical "notch" at 24kHz! In fact, if you look at the FFT in realtime when playing the song, you basically see that everything from 24kHz up is a mirror reflection of the material from 24kHz down! The only difference is that by about 40kHz, there's a roll-off in levels...

Bottom line is that this is not a real 96kHz recording. It's really 48kHz with some strange upsampling that creates a mirror image which fills up the upper frequencies making it look like there's something up there... I've actually seen this kind of thing before strangely enough back in the day with some early DVD-A releases. Anyone know what upsampling package does this kind of thing!? Or was it done purposefully!?

By the way, the album is analysed as DR7 using the foobar Dynamic Range Meter. There's no point IMO keeping this as 24-bits with so much dynamic compression... Resampling and dithering to 16/48 is therefore completely fine for this release and would save you plenty of space on the storage device! Remember, as I discussed previously, I absolutely refuse to keep anything less than DR10 in 24-bits; not only is it difficult already telling the difference between 16 and 24-bits, but as far as I'm concerned, pointless when dynamic peaks are clipped by volume compression. Heck, save yourself some money and just buy the CD release for this album... At least you get a case, booklet, CD and can resell when done with it...

An example of how the frequency spectral display can easily demonstrate something unusual going on... And a reminder of caveat emptor when buying "high-resolution" material. Analyse and verify that you're getting what you're paying for!

Happy Canada Day to my fellow Canucks. Wishing the Americans a great 4th of July!

Enjoy the music.

1 comment:

  1. nice article.
    This kind of mirror image can be achieved by upsampling (inserting 0 value samples). With sox this can be achieved by
    flac -dcs | sox -t wav - -r 96000 -t wav - upsample 2 | flac -8 - -o output96k.flac

    This however pushes the original signal level by 6 dB per 2x upsampling. So ideally you would like to make the stream pass through a LPF and compensate the vol 2 (slightly less actually to prevent clipping)

    Like so :
    ... | sox -t wav -r 96000 -t wav - upsample 2 sinc -23.75k -n 32767 -a 180 vol 2 | ...