在業(yè)界有三套采樣率并存：
44.1kHz及其下采樣、上采樣：22.05kHz, 11.025kHz，88.2kHz, 176.4kHz
48kHz及其下采樣、上采樣：24kHz, 12kHz,96kHz, 192kHz
8kHz, 16kHz, 32kHz, 64kHz

人耳的聽覺是有限的，介于20Hz到20kHz。跟據(jù)Nyquist采樣定理，采樣頻率只要超過信號帶寬的2倍就不會產(chǎn)生混迭。在數(shù)字媒體領(lǐng)域，如音樂CD的規(guī)范，都是以44.1kHz作為標(biāo)準(zhǔn)采樣率的。因?yàn)?4.kHz大于20kHz的兩倍，所以實(shí)際上44.1kHz的采樣率是足夠用的。

但是現(xiàn)在普遍在工程中都是使用48kHz或者96kHz頻率錄音，只有在最終母帶處理時(shí)才會轉(zhuǎn)成44.1kHz的CD格式，這樣減少多次采樣率轉(zhuǎn)換造成的失真。

而在電腦領(lǐng)域，作為音頻硬件codec標(biāo)準(zhǔn)的AC97規(guī)范只規(guī)定了48kHz。這造成幾乎所有的輸入、輸出信號都要被重新采樣（專業(yè)術(shù)語叫采樣率轉(zhuǎn)換，即 SRC）。SRC一般都會造成音質(zhì)的損失，較簡單（即較差）的SRC算法會造成音質(zhì)明顯劣化。但這已經(jīng)是一個(gè)既成事實(shí)了。
 
既然44K夠了,那為什么還要用192KHZ來錄音?

首先，20kHz只是大多數(shù)人的聽覺門限，也就是說，人耳對于20kHz以上的聲音很不敏感。注意不敏感并不意味著完全無法感知。大多數(shù)樂器（特別是鋼琴和弦樂器）的樂音含有豐富的高次諧波，用音樂術(shù)語來說即所謂的上方泛音。截止頻率為22.05kHz的CD音頻，的確會給聽?wèi)T了真實(shí)樂器的人一種不自然的感覺，尤其在高頻部分，因?yàn)槟慰固亟刂诡l率造成更高頻率泛音的信號失真。

其次，數(shù)字錄音通常都需要進(jìn)行后處理。音頻處理會對信號產(chǎn)生進(jìn)一步的失真，包括信號畸變、頻譜混疊，等等。如果錄音時(shí)僅僅用44.1kHz對原始信號采樣，那么在后處理前還得進(jìn)行上采樣（up-sampling），對采樣頻率進(jìn)行擴(kuò)展。由于這種擴(kuò)展是“假”的，實(shí)際上并沒有更多有用的原始信號，并且上采樣算法的優(yōu)劣也會影響原錄音信號的失真，所以這個(gè)做法并不可取。因此，通常的做法是用更高的頻率進(jìn)行采樣。

而現(xiàn)在的完全專業(yè)數(shù)字錄音棚中，則不再按CD標(biāo)準(zhǔn)的規(guī)范錄音、混音以及母帶，而是優(yōu)先使用HD音頻規(guī)范。即：
采用24Bit 48KHz、24Bit 96KHz、24Bit 192KHz 三種規(guī)格進(jìn)行錄音，當(dāng)然，24Bit 48KHz是一些小的錄音棚使用，因?yàn)樗麄兊奶幚砥髻Y源有限。而大的錄音棚，都清一色的使用24Bit 96KHz和24Bit 192KHz 進(jìn)行錄音。
那么，這樣的錄音規(guī)范，有什么好處？
1.符合HD音頻標(biāo)準(zhǔn)，這也是將來的主流標(biāo)準(zhǔn)，制作出的成品，可以直接應(yīng)用于HDCD、DVD-Audio、藍(lán)光唱片、數(shù)字音樂下載業(yè)務(wù)、數(shù)字對媒體播放機(jī)業(yè)務(wù)。
2.完全照顧數(shù)字影視視頻業(yè)務(wù)，多聲道電影視頻都會采用HD音頻規(guī)范。包括移動(dòng)便攜數(shù)字視頻設(shè)備都用它。
3.完全照顧消費(fèi)性音頻播放業(yè)務(wù)，比如：因特爾HD-Audio音頻標(biāo)準(zhǔn)，AC97音頻編碼解碼，便攜MP3/mp4/電話/游戲機(jī)最高音頻質(zhì)量播放。
 
目前，專業(yè)錄音行業(yè)的最高質(zhì)量標(biāo)準(zhǔn)就是：24比特定點(diǎn)位深、192000Hz采樣頻率，簡稱“24Bit/192KHz?！薄．?dāng)然，將來這個(gè)標(biāo)準(zhǔn)依然會繼續(xù)提高，向32Bit 384KHz進(jìn)發(fā)也是可能的。
 
實(shí)際上，現(xiàn)在的CD唱片市場上賣的產(chǎn)品（正版），最低級別的通常都是HDCD唱片，你買唱片時(shí)都會發(fā)現(xiàn)基本上都是HDCD標(biāo)識，也就是一張激光唱片包含兩種音軌：普通CD音軌和HDCD音軌。其中CD音軌記錄16比特44.1KHz信號（這是這張唱片的兼容內(nèi)容，照顧早期的CD播放機(jī)），HDCD音軌則記錄24Bit 96KHz信號（這才是該唱片的主要內(nèi)容）。普通的CD播放機(jī)只能播放CD音軌信號，而HDCD音軌則需要HDCD播放機(jī)才能播放（實(shí)際上現(xiàn)在的絕大多數(shù) DVD播放機(jī)都能播放HDCD，而現(xiàn)在的電腦則更沒問題了。）
 
下面是摘自論壇的一些有趣的討論，大多數(shù)的人（跟我一樣）搞不清楚為什么要用高的采樣率，比如96kHz,甚至192kHz,認(rèn)為這是出于對高采樣率的迷信，為了迎合盲目的消費(fèi)者，認(rèn)為這么高的采樣率是浪費(fèi)，是沒有必要的。也有人試圖解釋這背后的理論性和技術(shù)性原因， 包括：1）雖然大多數(shù)人聽不到20kHz以上的聲音，但是不代表所以的人都聽不到；2）人耳對瞬時(shí)信號（高頻）相對于穩(wěn)態(tài)高頻而言更容易感知，瞬時(shí)高頻往往帶寬較寬，有時(shí)候其帶寬可以包括一部分人耳可以聽到的頻率。雖然低通濾波器可以保留這部分頻率，但是總歸不如全部保留下來好。
還有人從過采樣的角度來分析，認(rèn)為有減少噪聲的作用，還可以降低模擬前置濾波器和模擬后置濾波器的復(fù)雜度。我認(rèn)為，過采樣的采樣率一般情況下遠(yuǎn)遠(yuǎn)比192kHz要高，而且過采樣的高采樣率僅僅體現(xiàn)在采樣過程中，在信號處理和保存時(shí)將到了正常采樣率（比如44.1，48，96，192，etc），這本身是過采樣的問題，跟本主題討論的問題不太相關(guān)。

上面給出的理由還不全面，一旦發(fā)現(xiàn)合理的解釋我再行更新。


http://www./showmessage/96236/1.php
What's the use of a 192 kHz sample rate?
為什么要用192kHz的采樣率？192kHz采樣率比44.1kHz有什么優(yōu)勢？人耳不是不能夠聽得到192kHz采樣率的全頻率嗎？
Why does DVD-Audio use 192 kHz sample rate? What's the advantage over 44.1 kHz? Humans can't hear the full range of a 192 kHz sample rate?

I agree there are a small percentage of humans who can hear above 20 kHz. However, DVD-audio uses a sample-rate of 192 kHz which allows a maximum frequency of 96 kHz. There is no known case of any human being able to hear sounds nearly as high as 96 kHz. I can agree with 48 kHz
sample rate and even 96 kHz sample-rate [maybe], but 192 kHz is just stupid.

So what’s the justification for using 192 kHz? Is it just a total waste of bandwidth and energy? Any proof to the contrary?

Thanks,

Radium


Some objective reasons客觀原因:

1) It's used in professional audio to accommodate nonlinear processing such as dynamic range compression.  The gain changes imposed by dynamic range compression are mathematically equivalent to modulation. Modulation produces sidebands.  Those sidebands *may* alias, under the
right conditions.  There have been some published papers on this. Though this may justify 96 kHz, it probably doesn't justify 192.

2) There's some evidence that humans' ability to hear high frequency transient signals exceeds their ability to hear high frequency steady state signals.  Somebody said he can hear steady state tones up to about 16 kHz, but he can hear music that contains tone bursts consisting of 6 cycles of a windowed (appears to be Hann) sinusoid at frequencies up to beyond the limit of his hearing. Though he perceive them as clicks, he definitely hear them.  Again, though this may justify 96 kHz, it probably doesn't justify 192. 

3) Others less verifiable, such as the claimed audibility of pre-ring associated with linear or non-minimum phase anti-imaging filters. Basically the higher sampling rate may allow more options here.  If true, this *might* justify 192 kHz.  Maybe.

Some subjective reasons: 主觀原因:

1) Marketing.

2) Bragging Rights.

3) Superstition.

4) Magical Thinking.

5) Self-Delusion.

6) Mass Hypnosis.

Personally I think that something like 64 kHz (possibly 128 kHz for mastering) would have solved all of these problems adequately.

-- Greg

>2) There's some evidence that humans' ability to hear high frequency
>transient signals exceeds their ability to hear high frequency steady
>state signals. 

I forgot to mention:  This can probably be fully explained by the fact that transient signals have a nonzero bandwidth associated with them. If that bandwidth extends down into the audible range, then, of course, the signal becomes audible.  From that standpoint, filtering-out the portion of the transient that is above the audible range should not affect the audibility of what is left.  Still, from a waveform fidelity point of view it might be beneficial to keep more of it.

Greg


I think there is great merit in sampling at 192k/s. These days a 192k/s 24 bit stereo DAC offers excellent noise and distortion specs. Such a high sample rate really makes the analogue filtering a lot easier. A 192k/s ADC is not much more expensive (a difference probably driven more by volume than complexity).

Actually transmitting and storing such sample rates makes no sense at all. 44.1k/s was a bit marginal, when you allow for the impracticality of the filters getting really close to 0.5fs. However, 48k/s should be good enough for any practical purpose.

For people who say supersonic sound can't play a part in a listening experience, trying being in a room with a high intensity of supersonic energy. Under some conditions (I'm not clear which) you can sense it, even though you can't hear it. It actually feels like something loud that you can't hear is going on. It’s a very odd feeling. That said, I've never found any evidence that this plays a part in any musical
experience. I see no reason to try to capture that energy in a recording, unless you feel your dog should enjoy a greater musical experience.

Steve



> Why does DVD-Audio use 192 kHz sample rate? What's the advantage over 44.1
> kHz? Humans can't hear the full range of a 192 kHz sample rate?

I think that the answer is aliasing avoidance. Take it this way:

- The audio band pass is limited to 16KHz, say 20KHz to get some extra marging for the most perfect ears on earth.

- As far as I know ANY audio digitization circuit uses a low pass filter at around 20KHz, so even a 192Ksps ADC or DAC will be band pass limited to 20KHz signals, as there is absolutely no need to manage audio signals with a higher frequency.

- If you use 44Ksps then you must insure that there is no power above 44/2=22KHz thanks to M. Nyquist, so your low pass filter must have a very sharp transition. As the filter will never be perfect you will get aliases. For example even if you use a 12th order filter (already difficult and
expensive to build) then the attenuation will be 'only' 72dB/octave, meaning that a 16KHz low pass filter will have an attenuation of only 50dB or so at 22KHz. And 50dB is not enough for good listeners as a -50dBc 'noise' is clearly audible.

- However if you use a 192Kbps sampling rate then the required performances on the low pass filter are drastically relaxed. This filter can keep a corner frequency at 16 or 20KHz, but even a 6th order filter will provide a at 86dB attenuation at 192/2=96KHz...

And as a 192Ksps sampling rate is far cheaper to build than a very very good low pass filter... That's the beauty of oversampling...

Does it make sense ?

Cheers,
Robert Lacoste
www.alciom.com
The mixed signal experts



>Does it make sense ?
>
>Cheers,
>Robert Lacoste
>www.alciom.com
>The mixed signal experts
>

Not a lot. As far as I'm aware there are NO ADCs that sample at the data rate of the output signal. For example the 44.1ksps ADC in my PC samples at 2.8224MHz. When you sample at that rate it is  trivially easy to make a gently sloping analogue lowpass filter that guarantees a lack of alias products. All further filtering and decimation is done digitally, where it is easy. THAT is what oversampling is all about, not using a 192ksps sampling rate.

d
--
Pearce Consulting
http://www.pearce.


Oversampled conversion does not require one to *store* information at the oversampled rate.

--
Oli


>Oversampled conversion does not require one to *store* information at
>the oversampled rate.

Fully right, but it is a low cost solution if you want to avoid the cost of a digital low pass filter & decimator...

Robert


> Fully right, but it is a low cost solution if you want to avoid the cost of
> a digital low pass filter & decimator...

The cost is in any case low, and a little extra cost in the acquisition chain is amply repaid in reduced storage cost of all the copies.

Jerry
--


On May 5, 8:14 am, rajesh <getrajes...@gmail.com> wrote:
>
> I don't mean literal repetition.Take for example a sine wave
> of 10khz and sample it with 1 mhz. One does think of
> samples getting repeated are at least they are close.

okay, rajesh, i'll try to clarify (or muddy) the waters here:

for a single sine wave (if it is already established that this is what you're looking at - a single sine wave), three sufficiently close samples (and spaced by 44.1 kHz or 1 MHz sampling rate are both
sufficiently close) will contain all the information you need.  Of course if there are errors (like quantization errors) in those three samples, you'll get a different sine wave reconstructed.

so rather than looking at a single sine wave at 10 kHz, let's consider a general waveform, but with one restriction of generality: a waveform *bandlimited* to just under 10 kHz.  we will call that bandlimit,
'B'.  now, if you're sampling at 1 MHz, it wouldn't be precise to say that there are samples getting repeated, but it is true that there is redundancy.  that is 50 times oversampled because it only needs
samples once every 1/20th millisecond.    you don't have 49 copies of each necessary sample, but the 49 samples in-between each 50th sample can be constructed from only the knowledge of those samples spaced by 1/20th millisecond.  so there is a *sense* of repeated samples (in
that in both cases of repeated samples and this oversampled case, 49 outa 50 samples is redundant), but it is, in fact, not the case.

now we *do* know that oversampling does, in the virtually ideal case, reduce noise and this (along with noise-shaping) is one of the neat properties of sigma-delta conversion.  for an N-bit converter, the theoretical roundoff noise is

    roundoff noise energy = (1/12)*( (full scale)*2^(1-N) )^2

that energy must be the area under the curve of the noise power spectrum.  note that the sampling rate is not a function of that.  If the roundoff noise is truly random in nature (a bad assumption for
very small signals, but not so bad for signals closer to full scale), then we think that the noise is white or flat, from -Fs/2 to +Fs/2. so, if the area under that constant function is this constant

    area = (1/12)*( (full scale)*2^(1-N) )^2

and the width is (+Fs/2 - (-Fs/2)) = Fs, then the height is

     1/Fs * (1/12)*( (full scale)*2^(1-N) )^2

so, as Fs gets larger, the height if this noise level (the amount of noise per Hz) gets lower, and if you can sacrifice some of the spectrum above your bandlimit, you can filter out all of the noise
from your bandlimit to Fs/2 and the level of noise has been reduced by a factor of B/(Fs/2) which is the reciprocal of the oversampling factor, Fs/(2B).  this is not unheard of, but it comes into play when
there is a limit to the number of bit in the word, N.  if that is the case (you have a very fast converter with fewer bits) you can make meaningful samples with word width wider than the A/D converter.  but
you have to oversample by a factor of 4 to get one extra meaningful bit.  that's how the math works out.  (this is not assuming noise-shaping.)

now, in audio practice, in the studio they get super-high quality A/D converters with, say, around 24 bits and sampling at a much higher sampling rate, maybe 192 kHz.  this is for initial recording, mixing,
editing, effects, etc.  i don't disapprove of them throw extra bits at this whether the need is disputed or not.

but eventually they are going into a CD of 16 bit words, two channels, and a 44.1 kHz sample rate.  that's 1411200 bits per second flying out at you. (or a DVD or SACD with a lot more bits per second.)  now, if the sample rate is increased (thus increasing the bit rate), how are we gonna 'compare apples to apples'?  if the bit rate increase with the sample rate, and if your bit error rate measured as bits of error/second remains constant, of course it will sound better as you increase Fs.  if it's 1 MHz sampling vs. 20 kHz sampling and you drop one sample per second in both cases, the 1 MHz data will care a lot less (in fact, 50 times less) than the 20 kHz data.  you are knocking out a larger portion of the data in the 20 kHz case.

but what if the portion was the same?  what if, in the 1 MHz case, you lost 50 samples per second compared to the 20 kHz case where you lose 1 sample per second.  which is worse?  in a recording or transmission environment, which is the case?  can you expect an equally noisy channel to have fewer bit errors per bit of data for the high sampling rate case?  what would be the mechanism for that?

r b-j



Suppose you want to digitize audio in the frequency range up to 20kHz.

If you digitize with a sampling rate of 48000, you need to use a filter which stops all frequencies above 24000Hz, but allows through all frequencies below 20000Hz.

However, if you digitize with a sampling rate of 192000Hz, you need a filter which stops all frequencies above 96000Hz, but allows through all frequencies below 20000Hz.

This is, in principle, an easier filter to build; and it's possible that may result in the quality of the filtered analog (pre-digitization) audio being better for the 192000Hz sampling than it is for the 48000Hz sampling.

So I think there are two separate questions:

1) do devices sampling at 192000Hz result in better digital files than devices sampling at 48000Hz? To answer this question, you need to use different sampling devices, as well as different sampling rates.

2) if you sample at 192000Hz, does it sound better if you store and play back the data at 192000Hz compared with downsampling to 48000Hz and storing and playing back at 48000Hz? To answer ths question, only one device is needed.

I think in practice your comment addresses the second question, but not the first.

Tim
 
 
還有一些討論：
The Difference Betweeen 96khz & 192khz
http://www./forum/49110-6-difference-betweeen-96khz-192khz