[vorbis-dev] encoding lots of speech

Gregory Maxwell greg at linuxpower.cx
Mon Jul 16 11:38:21 PDT 2001



On Mon, Jul 16, 2001 at 12:20:16PM -0700, Kevin Marks wrote:
[snip]
> Based on earlier discussion here, 'noise gating' your speech first 
> (putting it through a filter so that there is complete digital 
> silence between words) would help Vorbis make it a lot smaller, as 
> otehrwise Vorbis will carefully preserve the nuances of your 
> background noise in the recording studio well below the threshold of 
> hearing.

It can't be over-stressed how important good gating is on such signals to
keep the bitrate low.  The psycho-acoustical algorithms in Vorbis are loudness
sensitive, as the masking curves differ for different volume levels. Because
of this, Vorbis makes the assumption that the loudest sound is going to be
no more then a fixed loud amount (130db?) and normalizes the audio before
performing the psychoacoustics (and then reverses this on decode). The
calculate of 'loudest' is performed per block and MAX()ed with an
exponential decay from the prior loudness measures.

This is a very good and important feature for general compression, and is
one of the reasons that Vorbis does so much better for pieces with wide
dynamics (like much classical music) then mp3 , however, for speech signals,
if you don't pre-gate the audio, the compressor will end up wasting large
amounts of bits trying to accurately represent background noise the the
listener will never hear.

When doing some testing a while ago, I think I found it to be a 20% savings
on a fairly quiet speech signal.

The long term solution is for Vorbis to have special support for speech.
Such support could potentially include encode side changes that would allow
Vorbis to behave more like a speech source coder, and thus have competitive
bitrates while still having the capability for music in a single format.
(Which should be possible given the high degree of flexibility in Vorbis'
design).
 

--- >8 ----
List archives:  http://www.xiph.org/archives/
Ogg project homepage: http://www.xiph.org/ogg/
To unsubscribe from this list, send a message to 'vorbis-dev-request at xiph.org'
containing only the word 'unsubscribe' in the body.  No subject is needed.
Unsubscribe messages sent to the list will be ignored/filtered.



More information about the Vorbis-dev mailing list