A noise cross PSD estimator based on improved minimum statistics method for two-microphone speech enhancement dedicated to a bilateral cochlear implant

Coherence based methods have been successfully applied to dual-microphone noise reduction systems. These techniques showed good results when noise signals on two microphones were uncorrelated, but their performance decreased with correlated noises. It could be improved when the cross power spectral density (CPSD) of received noises is available. In this paper, an improved minimum tracking (IMT) technique for noise CPSD estimation was proposed. The performance of this technique was compared to two other noise CPSD estimators based on voice activity detection (VAD) and minimum tracking (MT) approaches. Evaluation was performed at four signal-to-noise ratios (SNR) and two interfering noise source configurations. Results showed a superiority of the IMT approach in terms of low computing time and quality indicated by the perceptual evaluation of speech quality (PESQ) scores. Then, subjective listening tests were carried out with 50 normal hearing listeners using a specific bilateral cochlear implant (BCI) simulator and utilizing the French Lafon database corrupted by additional babble noise. Results obtained with the proposed technique were better than the two previously mentioned noise CPSD estimators.