X-Fi Hybrid Sample Rate Converter
Jun 27, 2005 at 7:46 PM Thread Starter Post #1 of 17

korben_dallas

New Head-Fier
Joined
Apr 17, 2005
Posts
30
Likes
0
icon10.gif


The X-Fi sample rate converters use a unique three-stage architecture, as illustrated in figure 4. The first stage precisely doubles the original sample rate in a particularly computationally efficient manner. The second stage uses a poly-phase FIR filter to produce a sample rate that is four times the desired final output sample rate. The third and last stage reduces the sample rate by a factor of four to produce the desired sample rate. As with the first stage, this process is particularly computationally efficient.

fig4.jpg


The first stage doubles the original sample rate to provide a new signal at a sample rate of 88.2 kHz. The poly-phase FIR second stage converts this signal to a sample rate of 192 kHz by operating at a sample rate conversion ratio of two times 48/44.1. The final stage reduces the sample rate to the desired output rate of 48 kHz.

fig5.jpg


The expressed goal of the X-Fi sample rate converters is to provide audio quality limited only by the inherent analog noise of the analog-to-digital converters, while using many series and parallel sample rate converters.

The measured THD+N performance of X-Fi converting 44.1 kHz to 48 kHz is shown in figure 6. A full level sine wave at nearly any frequency produces distortion below -136dB (as measured by an Audio Precision distortion analyzer). The pass band ripple for this example is &plusmn0.00025dB.

fig6.jpg


The X-Fi sample rate converter supports a wide range of pitch shifts from zero to three octaves. The flexible audio routing of the X-Fi architecture means that audio can be taken through two sample rate converters in series to extend the pitch shift range even more.

Even with the advanced X-Fi hybrid SRC architecture this quality level doesn't come easy. The special purpose sample rate converter engine in X-Fi performs over 7 billion arithmetic operations per second.

Based on a novel architecture, the Creative X-Fi sample rate converters provide unprecedented audio fidelity, producing highly accurate 3D virtual environments and flawless synthesized music. Because of their extremely low ripple and exceptional THD+N performance, they form transparent digital patch cords, eliminating the headaches of sample rate synchronization. The Creative X-Fi audio processor provides flexible signal routing and the capability to operate at different sample rates to allow "bit-accurate" recordings when possible. When sample rate conversion is needed the user can simply listen to and record audio in an intuitive and creative way, thereby embracing the benefits of sample rate conversion and simultaneously enjoying phenomenal sound quality!
 
Jun 27, 2005 at 8:23 PM Post #2 of 17
Quote:

Originally Posted by korben_dallas
When sample rate conversion is needed the user can simply listen to and record audio in an intuitive and creative way, thereby embracing the benefits of sample rate conversion and simultaneously enjoying phenomenal sound quality!


Wait, what benefits?

I seriously don't understand how it is so impossible for Creative to produce a gaming card that doesn't resample. What a bunch of wasted man-hours. Just do all the spatial and reverb processing at whatever the source rate is (which this ultra-powerful new chip should be able to handle anyway!) and call it a day.
 
Jun 27, 2005 at 8:25 PM Post #3 of 17
Quote:

Originally Posted by Emon
Upsampling doesn't have much effect on sound quality.


It doesn't? Upsampling/resampling was (/is) one of the main reasons why Creative's past consumer cards weren't recommended for hifi audio. Their previous implementation was just poor.
 
Jun 28, 2005 at 1:36 AM Post #4 of 17
Quote:

Originally Posted by breez
It doesn't? Upsampling/resampling was (/is) one of the main reasons why Creative's past consumer cards weren't recommended for hifi audio. Their previous implementation was just poor.


You can use software resampling. Resampling really isn't the issue its made out to be.
 
Jun 28, 2005 at 4:47 AM Post #5 of 17
Quote:

Originally Posted by dasbin
Wait, what benefits?

I seriously don't understand how it is so impossible for Creative to produce a gaming card that doesn't resample. What a bunch of wasted man-hours. Just do all the spatial and reverb processing at whatever the source rate is (which this ultra-powerful new chip should be able to handle anyway!) and call it a day.



Not possible. For games and interactive 3D audio in general, you may have to mix together and process sound that has different source rates. I'm not sure that there exists a reasonable DSP architecture that can simultaneously process and mix samples with different rates.

Considering that they have put so much effort into SRC, I'm also guessing its not feasible to simply mandate in the PC audio archicture that ALL source material must be at exactly the same sample rate.

Do we have hard confirmation that you definitely can't playback 44.1KHz recordings bit-accurate? Or are we assuming not because of all the big talk from Creative regarding SSRC?
 
Jun 28, 2005 at 9:00 AM Post #8 of 17
Quote:

Originally Posted by maarek99
You can use software resampling. Resampling really isn't the issue its made out to be.


Of course, but I was just addressing the claim that upsampling/resampling hasn't got much effect on sound quality.
 
Jun 28, 2005 at 11:26 AM Post #9 of 17
Quote:

Originally Posted by korben_dallas
The measured THD+N performance of X-Fi converting 44.1 kHz to 48 kHz is shown in figure 6. A full level sine wave at nearly any frequency produces distortion below -136dB (as measured by an Audio Precision distortion analyzer). The pass band ripple for this example is &plusmn0.00025dB.

fig6.jpg



Where's the sine signal indication? Still even if it is THD+noise there should be an indication of a sine wave distortion at a certain frequency as a test sample. If the sine wave had too low amplitude it would mean no THD tests since no 1st, 2nd and 3rd order harmonics would be produced.
It would also be helpful (without sounding like an expert, i am not) - if they would define their passband clearly, because the THD+N graph can't be used to define the passband frequencies (?) - although the bandpass ripple is excellent! (in that example, as they say).

EDIT: Very good link: http://www.appsig.com/papers/tn070/tn070.html
 
Jun 28, 2005 at 1:19 PM Post #10 of 17
I suspect it shows the THD+N level expressed in dB at the frequencies shown (10-20000Hz). The distortion and noise remains very low (inaudible) with all tested frequencies.
 
Jun 28, 2005 at 9:59 PM Post #11 of 17
Jun 29, 2005 at 7:28 PM Post #14 of 17
Quote:

Originally Posted by breez
Of course, but I was just addressing the claim that upsampling/resampling hasn't got much effect on sound quality.


Well, it hasn't had an effect in my experience, and I was only addressing the "upsampling is hi-fi bliss, buy Creative!" thing going on in the first post.

Also, I apologize to the original poster and anyone else who read my post before it was deleted - that was just unneccessary of me.
 
Jun 29, 2005 at 8:31 PM Post #15 of 17
You're misinterpretting the plot. What they did was pass in a full-scale digital sinusoid and measure the THD+N at that frequency. All the recorded THD+N were then plotted vs frequency. They are (mercifly) not showing all the intermediate plots (what you're asking about), just the pertinent information, the bottom line if you will. If the card really is as good as the plot indicates, there should be no need for software resampling or complaints about audible artifacts.

Quote:

Originally Posted by Daroid
Where's the sine signal indication? Still even if it is THD+noise there should be an indication of a sine wave distortion at a certain frequency as a test sample. If the sine wave had too low amplitude it would mean no THD tests since no 1st, 2nd and 3rd order harmonics would be produced.
It would also be helpful (without sounding like an expert, i am not) - if they would define their passband clearly, because the THD+N graph can't be used to define the passband frequencies (?) - although the bandpass ripple is excellent! (in that example, as they say).

EDIT: Very good link: http://www.appsig.com/papers/tn070/tn070.html



 

Users who are viewing this thread

Back
Top