Beschreibung: Demonstrated in various prior investigations, wideband PESQ and POLQA are not capable to predict speech quality of artificial speech bandwidth extension (ABE) test conditions accurately enough. Based on a multilingual subjective listening test, evaluating ABE solutions from six different institutions, we present a novel objective measure particularly being able to estimate the speech quality of ABE-processed speech signals. The proposed measure employs a variety of features for ABE-related distortion detection. A combination of perceptually and non-perceptually influenced features are used as input to a support-vector-machine-based high-performance statistical model for establishing a link to an accurate speech quality estimate. Evaluation of the measure employs the ITU-Trecommended metrics correlation and rank order. Testing the predictive power of the proposed measure is done in three crossvaldation
setups: Disjoint languages, disjoint speakers, and disjoint ABE solutions. In a comparison, the proposed measure outperforms wideband PESQ and POLQA both in correlation and rank order.

Beschreibung: Low bit rate vector quantization (VQ) is omnipresent in today’s media transmission. With IP-based transmission the situation is that source-coded bits are typically either lost/deleted as a whole frame/packet, or they are received correctly. Assuming correctly received VQ symbols we show how to exploit vector-to-vector (i.e., residual temporal) redundancy at the decoder side for an improved reconstruction. It turns out that a feedforward neural network is an effective means for predicting better reconstruction vectors at the receiver in a system-compatible fashion, gaining up to 1 dB SNR depending on signal correlation and bit rate.