Non-Intrusive Binaural Speech Intelligibility Prediction from Discrete Latent Representations

Non-intrusive speech intelligibility (SI) prediction from binaural signals is useful in many applications. However, most existing signal-based measures are designed to be applied single-channel signals. Measures specifically take into account the properties of signal often intrusive - characterised by requiring access a clean and typically rely on combining both channels before making predictions. This paper proposes non-intrusive SI measure that computes features input using combination vector quantization (VQ) contrastive predictive coding (CPC) methods. VQ-CPC feature extraction does not any model auditory system instead trained maximise mutual information between output features. The computed predicting function parameterized neural network. Two functions considered this paper. Both extractor simulated with isotropic noise. They tested real For all signals, ground truth scores (intrusive) deterministic STOI. Results presented terms correlations MSE demonstrate able capture relevant modelling outperform benchmarks even when evaluating data comprising different noise field types.