[Speex-dev] Speex for sampling freq >48KHz

Jean-Marc Valin Jean-Marc.Valin at USherbrooke.ca
Mon Mar 27 05:19:57 PST 2006

Just forget about using Speex as the basis of a lossless codec. All the
things that make CELP good (noise weighting, VQ, ...) are simply useless
for lossless. It will not get you anywhere in terms of improving
lossless coding. If you really want to do something in lossless, you
better post to the FLAC list and ask if there are promising things they
didn't use in FLAC because of some design decisions they made (e.g.
because of independent blocks). The only thing *I* can think of where
you could (theoretically) improve lossless coding for speech (only) is
to use a long-term predictor. 


Le lundi 27 mars 2006 à 05:01 -0800, devilal sharma a écrit :
> Hi,
> I chose speex initially because i had some work in VQ on speex i.e.
> modifying split VQ to GMM based parametric VQ and I thought If I train
> the GMM based VQ codebooks with audio signal and then do audio coding
> with speex, I probably get a better(smaller) residual signal even with
> speex. But I couldnt get that.
> I was trying to get a lossless bitstream by MUXing the speex-bitstream
> and the entropy encoded error signal from speex. 
> I still have 2 months time but I need to do some useful and
> fundamentally strong work(even if  its  a small improvement). I can
> work day and night, But want to come up with some results. Results
> could be anything which make sense.
> If you could suggest me some work in order to improve even a small
> part in lossless coding, I would be grateful.
> I am sorry for any act of not respecting you in the past. I consider
> you at the top in guiding me.
> Thanking you,
> Regards,
> Devilal
> Jean-Marc Valin <Jean-Marc.Valin at USherbrooke.ca> wrote:
>         > I have one doubt again, that is Vorbis use DCT/MDCT based
>         algorithm
>         > and also use psychoacoustic model so this is lossy codec. 
>         Speex is also a lossy codec.
>         > And I dont think it ca regenerate a better matching waveform
>         than
>         > speex. 
>         At bit-rates above 32 kbps, Vorbis tends to produce better
>         results than
>         Speex, even for speech. The only advantages of Speex over
>         Vorbis at
>         these high rates is the lower latency and lower encoding
>         complexity.
>         > Then there comes FLAC which is the perfect answer to my
>         question, I
>         > suppose. But my concern is this that FLAC use simple
>         prediction
>         > algorithm and doesnt use any CELP based algo which could
>         have model
>         > the waveform coding by having a large codebook and comparing
>         the
>         > residual signal and selecting the codebook index.
>         FLAC is entirely different. You need to choose between perfect
>         quality
>         and low bit-rate. FLAC isn't better or worse than Speex. If
>         you can't
>         decide between the two, then you've obviously have no idea
>         what your
>         after in the first place.
>         > For this, shall I start understanding and modifying FLAC
>         itself in
>         > case I need to do something for lossless coding or I can try
>         on Speex
>         > and than apply entropy coding.
>         > I am getting quite good(comparable) results for audio
>         signal(44.1KHz)
>         > if use speex and separate entropy coding.
>         You apply entropy coding to the Speex bit-stream or you encode
>         the Speex
>         error to make a lossless codec (which I already mentioned is a
>         stupid
>         idea)?
>         > Please suggest me clearly as I have very small time left to
>         wrap up my
>         > work to submit.
>         To be honest, that's the least of my problem. If you were
>         clear in the
>         first place and listened to advice I already gave, you might
>         have been
>         better off already.
>         Jean-Marc
> ***************************************************************************
> Devi Lal Sharma
> Final Yr. Dual Degree(Communications)
> Mobile: +91-9986423985(I am currently in bangalore)
> email: devilal_sharma at yahoo.com, devilal at gmail.com
> ***************************************************************************
> ______________________________________________________________________
> Yahoo! Messenger with Voice. PC-to-Phone calls for ridiculously low
> rates.

More information about the Speex-dev mailing list