Quantization in voice classification over remote system by hidden markov model

Abstract EN

Automatic Voice Classification system over web deserved special interest for the many possible applications such as: web navigation, data entry, data base access, verification, voice and natural language understanding systems and for general enhancement of the internet experience.

Combining AVC systems and network environments presents the concept "Distributed Voice Classification", which, has added extra complexity to both: the classification process and the transmission process.

For the classification process , due to the lack of speed on the web transfer protocols , lower bit rates have to be used in the transmission process so as to keep the AVC system working in real time as it is intended to be ,which in turn causes the AVC classification accuracy to degrade .

The compromise is to lower the bit rate and keep the classification accuracy unchanged.

In distributed voice classification, low complexity clients acquire voice and transmit it to a remote server for classification.

So that compression by quantization is required before data is transmitted and the computation at the client should be kept at a minimum level.

In our work, the design of optimal quantization for distributed classification applications is considered in the context of a voice.

The obtained results indicate that new distortion measure is a better metric (when compared to mean square error) for optimizing encoders used in distributed classification applications.