So, I would like to know how are the data vectors that come out of the
encoder composed, by that I mean, how are the 4 parameters, pitch, energy,
voicing and spectrum organised in the bitstream. I would like to extract
the parameters in order to work on them.
On Fri, Nov 3, 2017 at 12:22 PM, Matei Alexandru Coltoiu <
coltoiu.ma...@gmail.com> wrote:
> David, the current block scheme of soft-decision demodulators (as far as
> I konow) includes a demodulator that still outputs integer values, but it
> outputs current quantisation tresholds and accepts a feedback for the
> decision making process. My theoretical demodulator would be a non-decision
> demodulator which outputs raw data. It is true that you would have to have
> much more memory and processing power available in order to process such
> large data quantities but today we have such posibilities.
> It is quite intersting that already in 1993 an article was published,
> Dynamic Adaptation of Quantisation Thresholds for Soft-Decision Viterbi
> Decoding with a Reinforcement Learning Neural Network. It uses a soft
> demodulator wich uses feedback from a neural network. That's a concept 25
> Years old, when neural networks were more like a concept with little
> practical implementations. Today we already have such networks available.
> So there are 3 steps we could take: Develop a non decision making
> demodulator, feeding raw data into a neural network, feeding processed raw
> data intro the modified decoder, made able to process floating point binary
> data.
>
> Tomas, the encoding side is a different story. The H 2.64 loop filter
> prevents some encoding artefacts from becoming to obvious. Now, first we
> would have to consider wich artefacts we would like to smoothen, encoding
> artafacts or specific possible BER artefacts. Then, we would have objective
> and subjective artefacts, that means you would have artefacts affecting the
> objective form of the signal, but we do not know unless we test with a
> group of subjects how muh these artefacts affect the subjective perception
> of the speech.
> The only simple approach for an objective filtering, for both encoding and
> BER artefacts I see here is a filter which is modelled by a neural network.
> You take the filter parameters and model them with the help of a neural
> network which compares original and encoded signal.
> However, I would leave this for another topic and I would concentrate on
> the receiver part for now.
>
--
Colțoiu Matei Alexandru, telefon: 0373741128 <0373%20741%20128>
whatsapp 0748021325 <0748%20021%20325>/0770663003 <0770%20663%20003>.
Skype: rav0rmat
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Freetel-codec2 mailing list
Freetel-codec2@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/freetel-codec2
