This is a joint research program led by the Cinvestav in Mexico and the ENST de Bretagne in France.
It is effectively led as a scientific cooperation exchange between France and Mexico, managed by ECOS-Nord in France and ANUIES/CONACYT in Mexico: while the responsible teachers in the respective universities manage the project, a French student goes to Mexico for four months from April 20th to August 20th, and a corresponding Mexican student will go to the French university from September on.
A digital communication system is typically made of two fundamental elements, with opposing functions: source encoder does compress information by suppressing redundancy in the signal, allowing, while keeping a low deterioration of the signal quality, to reduce bit rate, and thus to increase the capacity of the transmission medium. However, this natural redundancy is naturally robust w/r to noise. To compensate for the reduction of this robustness, the channel encoder, whose role is to fight against transmission errors, artificially introduces redundancy, to allow for error correction.
Until now, the two processes have been optimised separately. We are now looking to optimise globally the source and channel encoders, in order to better quality for a given bit rate.
The current speech joint source-channel encoding techniques assume errors caused by a symmetrical binary channel, as well as isolated, and as such easy to correct. We know how to compute the optimised structure of the vectorial quantifier for a CELP-type speech encoder, which is designed for the symmetrical binary channel. This is only a first step in joint optimisation for in reality, channels are less ideal: they could show gaussian perturbations, as observed with satellite communications, or perturbation following a Rayleigh law, as observed with multipath communications, such as mobile communications. In those cases, errors are grouped, and as such more difficult to correct.
The objective of the project is thus to progress in the elaboration of algorithms by choosing more realistic cases. For image encoders, solutions were designed that associate a DCT and vectorial quantification source encoding with a QAM modulation. Though the transmitted signals have different qualities, the student might want to get inspiration from these ideas for speech encoding. The student will as well consider extending the optimisation of vectorial quantisation with gaussian or Rayleigh noise.
More specifically, the placement shall include the following steps: