Matching pursuit for channel selection in cochlear implants based on an intelligibility metric

In this paper the earlier proposed short-time objective intelligibility predictor (STOI) is simplified such that it can be expressed as a weighted ℓ₂ norm in the auditory domain. Due to the mathematical properties of a norm, STOI can now be used with the matching pursuit algorithm in the n-of-m channel selection technique as found in several cochlear implant (CI) coding strategies. With this technique only a subset of frequency channels (electrodes) are stimulated, such that important channels can be updated more frequently and less significant channels are omitted. Intelligibility predictions with acoustic CI-simulations for normal-hearing listeners indicate that more intelligible speech is obtained with the proposed method compared to a conventional channel selection method based on peak picking. Reasons for this difference in performance are: (1) STOI considers an analysis window of a few hundreds of milliseconds in order to account for important low temporal modulations for speech intelligibility and (2) spectral leakage per channel is accounted for in the mathematical optimization process.