Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.compression > #2206 > unrolled thread

low bitrate speech codec enquiry

Back to article view | Back to comp.compression

Contents

  low bitrate speech codec enquiry Nimo <azeez541@gmail.com> - 2014-02-05 07:29 -0800
    Re: low bitrate speech codec enquiry glen herrmannsfeldt <gah@ugcs.caltech.edu> - 2014-02-05 17:06 +0000
    Re: low bitrate speech codec enquiry BGB <cr88192@hotmail.com> - 2014-02-05 14:40 -0600

#2206 — low bitrate speech codec enquiry

Hi there,
I'm back after a very long time.
I googled for low bitrate speech codecs for information on
what is the minimum bitrate a speech codec offers and I found 
www.speex.org which offers sound at a mimimum of 2500 bits per sec

I wanna know, Are there any other codecs which offer speech less than 2.5kbps? what is the minimum bitrate one can expect for the speech?
I heard that some vocoders can play sound at 600 bits per second

Thanks for sharing the info.

    so long,
      nimo.

[toc] | [next] | [standalone]

#2208

Nimo <azeez541@gmail.com> wrote:

> I googled for low bitrate speech codecs for information on
> what is the minimum bitrate a speech codec offers and I found 
> www.speex.org which offers sound at a mimimum of 2500 bits per sec
 
> I wanna know, Are there any other codecs which offer speech 
> less than 2.5kbps? what is the minimum bitrate one can 
> expect for the speech?

Seems to me that the lower you go, the less the output will sound
like the input. The limit should be a speech to text on one end,
and text to speech on the other, where words are sent in between.
(Presumably appropriately coded to minimize the bits). Just before
that, decode and send the phonemes along with appropriate timing.

You might also factor out some of the spectral information, what some
might call voice-print, and only send that at the beginning. No need
to keep sending things that don't change.

> I heard that some vocoders can play sound at 600 bits per second

-- glen

[toc] | [prev] | [next] | [standalone]

#2211

On 2/5/2014 9:29 AM, Nimo wrote:
> Hi there,
> I'm back after a very long time.
> I googled for low bitrate speech codecs for information on
> what is the minimum bitrate a speech codec offers and I found
> www.speex.org which offers sound at a mimimum of 2500 bits per sec
>
> I wanna know, Are there any other codecs which offer speech less than 2.5kbps? what is the minimum bitrate one can expect for the speech?
> I heard that some vocoders can play sound at 600 bits per second
>
> Thanks for sharing the info.
>

while not a direct answer, I have made some observations:

intelligible speech can be generated from raw 1 bit/sample audio, and 
some filtering can improve the sound-quality, as well as using an 8kHz 
sample-rate and similar;
the output in-turn seems subject to vector-quantization (approximating 
the bit-patterns with a smaller alphabet via a lookup table), and likely 
LZ77 and arithmetic coding could also be used.

in what experiments I had done (at lest with the 1 bit/sample and VQ 
stuff), the sound mostly just comes out a bit "tinny".


for example, a possible 8->3 bit reduction table (just sort of pulled 
out of the air):
000 <-> 00000000
001 <-> 11111111
010 <-> 00001111
011 <-> 11110000
100 <-> 00110011
101 <-> 11001100
110 <-> 01010101
111 <-> 10101010

then you can cram 32 samples into 12 bits, or use 16-bits with the other 
4 bits as an escape case or interpolant selector or similar.

or, maybe use 4 bits and encode more patterns, or a larger group to 
exploit more possible redundancy, ...


to a lesser extent, "similar" is done with my BTAC codec, albeit it more 
uses bits for selecting between upper/lower range values (sort of like 
DXTn, as opposed to encoding samples directly), but this was mostly for 
trying to get better sound quality than ADPCM for a lower bitrate and 
for fixed-size sample blocks.

in stereo mode, it is around 1 or 2 bits/sample, so 88kbps or 176kbps 
for 44.1kHz audio, or ~ 16/32kbps for 8kHz. BTAC-encoded audio also 
seems subject to further compression via Deflate, but this isn't really 
used much at present (it is mostly used for random-access, with any 
Deflate compression mostly due to packaging inside a ZIP file or similar).


I am not sure the lower-limit here, as I haven't really tried using the 
raw 1-bit audio scheme for audio compression (it was more done in terms 
of signal processing, mostly some fiddling with trying to synthesize 
"intelligible" speech mostly from bit-patterns and bitwise arithmetic, 
and for trying to isolate bit-patterns for various phonemes).

it seemed promising, though the audio quality was a bit worse than with 
a straight diphone synthesizer, so I don't know. could still be worth 
further investigation though.


or such...


>      so long,
>        nimo.
>

[toc] | [prev] | [standalone]

Back to top | Article view | comp.compression

csiph-web