🚀 go-pugleaf

There is no way you can get an hour of speech into a megabyte
with reasonable quality if you just digitize waveforms. Suppose
you sample at a resolution of 8 bits, which is the resolution of
the cheapo ADDACs you can buy for PCs and Macs. For research
purposes and hi-fi people use higher resolution.  That means your
megabyte gets you 1024000 samples. At 3600 seconds in an hour,
that means 284.44 samples per second is the maximum sampling rate
you can use. The corresponding Nyquist frequency is 142.22 Hz,
meaning that you can only represent frequencies below this level.
This is WAY too low. For music people use sampling rates around
44K samples/sec in order to get frequencies up to over 20KHz. For
speech you don't need anything that high. For speech research we
typically sample at 20K samples/sec and low pass filter at 8KHz.
That covers everything significant for speech.  For some purposes
we sample at 10KHz with low-pass filtering at 4KHz.  Engineers
often sample at 8K samples/sec because they expect to be working
with telephone speech, which is limited to the region below about
3200 Hz. Already we're talking about degraded, though
intelligible, speech. So you can see that if you just want to
record and edit waveforms, there is no way you can cram an hour
of speech into a megabyte.  To store speech at around 2000
bits/second as you wish to do is possible but requires non-trivial
coding, and I'm not sure that you will like the quality that
results. I think you're going to need a bigger disk.

If you want anything that sounds even like a cheap walkman (in mono)
you'll need way more than a megabyte for 1 hour.  1 meg may work for
about 1.5 to 2 minutes with almost cheap walkman sound.  If you can
prove me far wrong on this, please do, as I'd be interested in
discovering a way to do such a thing!

    Alexander Morris           ... Be excellent to each other,
    Carnegie Mellon                my excellent friends ...

All right, my goal of an hour of sound in a megabyte betrayed my
ignorance of signal theory.  So let's bump up the space requirement by
an order of magnitude.  Are we into the real world yet?
-- 

ergo@netcom.uucp			Isaac Rabinovitch
netcom!ergo@apple.com			Silicon Valley, CA
{apple,amdahl,claris}!netcom!ergo

WISE SAYING NEEDED.  Must reflect positive human values.  Gentle humor a
...  
Hey, it was a JOKE!

In article <16002@netcom.UUCP> ergo@netcom.UUCP (Isaac Rabinovitch) writes:
>All right, my goal of an hour of sound in a megabyte betrayed my
>ignorance of signal theory.  So let's bump up the space requirement by
>an order of magnitude.  Are we into the real world yet?

Well, if you just want to use simple sample and not any kind of coding,
you'll need a bit more. Here's how to do the calculation. Suppose you're
going to use one byte per sample. That gives you passable resolution,
though not what you'd want for research or real hi-fi. It's also the
resolution of all of the cheap digitizers. To get good quality speech
you should sample at a minimum of 12K samples/second. This gives you
room for anti-aliasing filtering at about 5KHz (you need to allow
for the fact that the filter cutoff is not perfectly sharp). So,
you need 12K bytes per second. That is 1.2e4 * 3.6e3 = 4.32e7
bytes per hour, or 43.2MB per hour. So you need another half an order
of magnitude to get into the ballpark.

Bill

In <16150@csli.Stanford.EDU> poser@csli.Stanford.EDU (Bill Poser) writes:

>
>.... Here's how to do the calculation. Suppose you're
>going to use one byte per sample. That gives you passable resolution,
>though not what you'd want for research or real hi-fi. It's also the
>resolution of all of the cheap digitizers. To get good quality speech
>you should sample at a minimum of 12K samples/second. This gives you
>room for anti-aliasing filtering at about 5KHz (you need to allow
>for the fact that the filter cutoff is not perfectly sharp). So,
>you need 12K bytes per second. That is 1.2e4 * 3.6e3 = 4.32e7
>bytes per hour, or 43.2MB per hour. So you need another half an order
>of magnitude to get into the ballpark.
Your figure is a *little* off -- a K is 1024, not 1,000.  (Computers
are so dumb they have to count on their fingers -- both of them.)  But
I take your point -- casual digital recording would seem to be beyond
the current generation of computer media, and probably the next as
well!  And by the time 43 meg mass storage has joined 64K RAM in the
netherworld of "used to think that was a *lot*", we'll have more
interesting problems -- or more drastic ones.
-- 

ergo@netcom.uucp			Isaac Rabinovitch
netcom!ergo@apple.com			Silicon Valley, CA
{apple,amdahl,claris}!netcom!ergo

WISE SAYING NEEDED.  Must reflect positive human values.  Gentle humor a
...  
Hey, it was a JOKE!

In article <16015@netcom.UUCP> ergo@netcom.UUCP (Isaac Rabinovitch) writes:
>Your figure is a *little* off -- a K is 1024, not 1,000.

I know. But a 2.4% difference isn't much when doing back-of-the-envelope
calculations.

In article <16150@csli.Stanford.EDU> poser@csli.stanford.edu (Bill Poser) writes:
>In article <16002@netcom.UUCP> ergo@netcom.UUCP (Isaac Rabinovitch) writes:
>>All right, my goal of an hour of sound in a megabyte betrayed my
>>ignorance of signal theory.  So let's bump up the space requirement by
>>an order of magnitude.  Are we into the real world yet?
>
>Well, if you just want to use simple sample and not any kind of coding,
>you'll need a bit more. Here's how to do the calculation. Suppose you're
>going to use one byte per sample. That gives you passable resolution,
>though not what you'd want for research or real hi-fi. It's also the
>resolution of all of the cheap digitizers. To get good quality speech
>you should sample at a minimum of 12K samples/second. This gives you
>room for anti-aliasing filtering at about 5KHz (you need to allow
>for the fact that the filter cutoff is not perfectly sharp). So,
>you need 12K bytes per second. That is 1.2e4 * 3.6e3 = 4.32e7
>bytes per hour, or 43.2MB per hour. So you need another half an order
>of magnitude to get into the ballpark.

Well, to furthur draw this out, let me get my 2 cents worth in here.
If you really wanted to be minamalist, you can get reasonable speech
sounds at about 6Khz to 8Khz.  (6 sounds a lot like a telephone, 8 is
more like a Cheap AM radio).  Assuming you also want "numerically
perfect" (ie. no error) reporduction, you can almost get 2 to 1 
compression using some form of delta encoding (differences between
samples).  If you are digitizing pure speech (no background sound or
music) you also will find that about 17% of this can be stored as silence.
(Again, this is from one of my samples, your mileage may vary).  If you 
care to go to approximate compression, you can likely get as much a 3
or 4 times compression.

Anyway, time for some math:

Say, sampling rate of 8KHz, 2 to 1 compression, and silence is encoded
as a 3 byte command so its contribution to a overall size of the file
will be negligable if we assume that silence is no change in the input
sample for say 300 samples (.0375 seconds).

8000 samples/sec * 3600 sec = 28.8 x 10^6 samples
2 to 1 compression gives      14.4 x 10^6 compressed samples
17% of the above is roughly (very roughly) 2.4 x 10^6

So your total number of bytes is 12 x10^6

Is that any closer ???
                                        Richard Snider
------------------------------------------------------------------------
Where: ..uunet!mnetor!yunexus!xrtll!rsnider    Also:  rsnider@xrtll.UUCP
An unbreakable tool is useful for breaking other tools.