[jefemeister]

With my recent purchases of an Emu 0404 I thought it would be interesting to capture the digital data coming from a few sources and do direct file comparisons between them. I used a Pioneer Elite DV-47A and Toshiba 3960, and a super-crappy freebie video cable as interconnect.

The general procedure:

1. Setup PatchMix for Digital input, external clock, with a send to Wave L/R.
2. Open WaveLab Lite (I can't get Cubasis to work for the life of me), cue up the CD Player, and record.
3. Find the very first non-zero sample in the recording and trim the beginning of the file to this point. Then find a suitable time location to trim the end of the file.
4. Open the wav in GoldWave (can be downloaded free) and Save As a text file with "integer float" data representation.
5. Use a file comparison program (like Examine Diff, also free) to compare the results. Notepad is also useful for this provided the files aren't too big.

I did this for my two players a bunch of times to get repeatability results. I also recorded with the 0404 set to use its internal clock. Reference file was a trimmed EAC rip.

The results:

I'm still working on this project so expect more in the future. I limited myself to working with 10 second files, I obviously would like to use larger ones. I used a pristine CD of "Byrd in Hand" for my tests. I would also like to try more complex music (I'm thinking Aphex Twin) and some different degrees of scratched discs.

With the 0404 set to external clock, the Pioneer matches the EAC rip 100% perfectly over 5 measurements.

With the 0404 set to external clock, the Toshiba never matches the EAC rip at all although the data does track the EAC file in character. Ie, it looks like a Byrd, and sounds like a Byrd, but it isn't a Byrd. The Toshiba recordings all match each other 100% though, indicating that the Toshiba is doing the same thing wrong every time. I have a suspicison that there is some Asychronous sample rate conversion going on.

With the 0404 set to internal clock, both the Pioneer and Toshiba files are probably 99% identical to the respective references. There are a few sample differences spread throughout the files which tend to be no more than 2-3 samples in a row. I wouldn't use the internal clock unless you have to.

Hope someone finds this interesting, I still have some investigation to do. Someone else should run some tests too with their players.

 

[Wodgy]

This is very interesting. Even ignoring jitter, transports clearly aren't the same after all.

Can you post a short excerpt from the Pioneer versus the Toshiba files, so we can have a look at how different they are? Also, can you subtract the two waveforms in an audio editor and take an FFT of the result, so we can see the distortion spectra?

 

[jefemeister]

Quote:

Originally Posted by Wodgy This is very interesting. Even ignoring jitter, transports clearly aren't the same after all. Can you post a short excerpt from the Pioneer versus the Toshiba files, so we can have a look at how different they are? Also, can you subtract the two waveforms in an audio editor and take an FFT of the result, so we can see the distortion spectra?

Once I get Matlab running on my box (having problems with FlexLM), I definitely plan on doing these types of tests and producing a semi-formal paper on it. I will post example Toshiba/Pioneer and Internal/External data as soon as I get a chance.

 

[Mr.Radar]

Have you tried playing with the "PCM downsampling" or "Digital Output" options under the Audio setup on the Toshiba?

 

[jefemeister]

I have done a little more experimentation:

With 0404 set to use the external clock and using a brand-new CD, I recorded 1 minute of data from both the Pioneer and Toshiba, 3 times each. These were compared to each other as well as to a perfect EAC rip.

Comparison of Pioneer to Pioneer: 100% exact.
Comparison of Pioneer to EAC: 100% exact.
Comparison of Toshiba to Toshiba: 100% exact.

Comparison of Toshiba to EAC:
.....mean of difference = 0.85 quantization levels.
.....standard deviaiton of difference = 824.13 quantization levels
.....percentage of samples equivalent = 0.18%
.....for comparison, mean of just EAC = -.6*QL, std. dev = 4259*QL
.....I made sure that the difference did not arise from a time offset or gain.

The following figure is charted data for Toshiba vs. EAC, includes a FFT:



This FFT plot may be a little more helpful than the above one:


Click on either picture for a larger version.

There are more data and experiements on the way. . .

Mr. Radar, I'll go back and check this, but the Toshiba should default to redbook pcm upon detecting a CD. I looked through the menu when I first bought the unit and do not remember there being any downsampling/output options.

 

[Dane]

Keep up the good work! One of the most interesting threads I've seen. I never imagined that bits could actually get lost - I guess the Toshiba must be really bad.

 

[gaboo]

Hmm, maybe the phat Toshiba chip in 39[56]0 does resample to 48/96/192Khz. After all it is a DVD player. Most likely, they want to take advantage of the 192Khz DAC, so it makes sense to do this before sending data to the onboard DAC. The question is where is the digital out hooked.

Anyway ASRC in consumer stuff has come a long way since the 10K1.

 

[Canman]

Cool stuff Jeff

Are you sure that both units have equivalent sample rates and word lengths at the digital output?

If there is asynchronous sample rate conversion going on in the Toshiba, I suppose that would explain the difference in samples since new samples would be estimated by formula. The question I have is, how significant of a difference is one quantization level when there are 65,536 levels at 16 bit?

 

[Glassman]

there's no practical advantage resampling to higher rates other than you can use single clock generator with 'dvd' samplerate multiples, but apparently the redbook s/pdif output of toshiba is 44.1.. am I right Jeff? is it 16bit or 24bit?

 

[jefemeister]

Quote:

Originally Posted by Canman Cool stuff Jeff Are you sure that both units have equivalent sample rates and word lengths at the digital output? If there is asynchronous sample rate conversion going on in the Toshiba, I suppose that would explain the difference in samples since new samples would be estimated by formula. The question I have is, how significant of a difference is one quantization level when there are 65,536 levels at 16 bit?

I'm being careful not to make any claims about sound quality just yet. I have a feeling that the 2nd FFT I posted above is heavily skewed due to the lack of high-freq energy in the track. I am going to make recordings tonight using pure whitenoise as input. Also will try a pure 1kHz tone to get a better idea of harmonic content.

I am almost positive that this is the work of an ASRC. Note that not all SRC is for up/downsmampling. It is also used in a 1:1 ratio purely to synchronize two different clocks (input stream and output stream). I am confused as to why it's also being fed to the digital out though. I still have to double-check that the Toshiba is outputting 44.1 and not 48. If it really is outputting 48 than a lot of this thread will be retracted.

A difference of one QL is 20*log(1/2^16) which is -96dB. The issue here is that we're dealing with an average of 825 QL's which is equal to -38dB. Wordlength will not matter.

 

[jefemeister]

Quote:

Originally Posted by Glassman there's no practical advantage resampling to higher rates other than you can use single clock generator with 'dvd' samplerate multiples, but apparently the redbook s/pdif output of toshiba is 44.1.. am I right Jeff? is it 16bit or 24bit?

When I first bought the 3960 I cracked it open and only found a 27MHz clock which is for video. Usually there is a PLL that is used to generate 44.1, 48, etc. I do not know if the 3960 upsamples prior to conversion but it really doesn't matter here. Clasically, digital outputs should remain 44.1/16. It is the responsibility of the DAC to do more with it. Toshiba may just be converting everything to 24/48 from the start though. I didn't notice much of a servo board for the 3960 indicating that the audio processing and servo functions may be more tighly integrated than other designs. A lot will be answered once I verify the output rate of the 3960.

 

[jefemeister]

I have the update on the Toshiba's output rate. It is 44.1kHz and 16 bit. The only audio setting for the player is "raw" or "pcm" audio out. I get the same output regardless of setting, both of which are wrong. I therefore conclude, as I've long expected: The Toshiba 3960 is a piece of crap.

More results and hopefully some aditional players to come. . .

 

[Wodgy]

Can you post some short excerpts (i.e. lists of numbers) from the Toshiba output versus the Pioneer output so we can see what it is doing in the time domain?

 

[jefemeister]

Quote:

Originally Posted by Wodgy Can you post some short excerpts (i.e. lists of numbers) from the Toshiba output versus the Pioneer output so we can see what it is doing in the time domain?

Sorry, forgot you asked me that before. The following are 1 second mono sections from the Pioneer and Toshiba using external clock for the E-mu. This is a small subset of the data used for the above plots.

pioneer.txt
toshiba.txt

These were supposed to be mono files, but it looks like WaveLab merely copied the left channel to the right as opposed to truncating the right. That's why you see the same values repeated twice in a row. I'll fix it when I get a chance.

 

[Wodgy]

Thanks! Very interesting. Those are substantial errors. They're also not obviously simple or consistent (e.g. it's not the last two bits or so that get munged).

It does look like some kind of filtering has been done to the Toshiba's datastream. It's not unusual for transports to do 2x oversampling at the transport level these days. I wonder if such transports also have a running FIR filter as part of an error correction scheme.

Regardless, very distressing.