Non-intrusive bit-rate detection of coded speech
Dushyant Sharma, Uwe Jost, Patrick A. Naylor · 2017
We present a non-intrusive codec type and bit-rate detection algorithm that extracts a number of features from a decoded speech signal and models their statistics using a Deep Neural Network (DNN) classifier. We also present a method for reducing the computational complexity and improving the robustness of the algorithm by pruning features that have a low importance and high computational cost using a CART binary tree. The proposed method is tested on a database that includes additive noise and transcoding as well as a real voicemail database. We show that the proposed method has 25% lower complexity than the baseline, 19% higher accuracy in the bitrate detection task and 10% higher accuracy in the CODEC classification experiment.