Speech Feature Fidelity from a Generative Auditory Transduction Model

Speech production and auditory transduction evolved together. Auditory transduction efficiently encodes sound pressure waves and recently has been emulated using a generative deep learning model, “WavIHC”. WavIHC runs in real-time, and can approximate the outputs of a state-of-the-art phenomenological auditory transduction model. This paper seeks to understand how well speech information is retained by WavIHC's approximations. By evaluation with two statistical metrics, Signal-to-Distortion Ratio (SDR) and the Neurogram Simiarlity Index Measure (NSIM), we find that WavIHC more closely approximates the phenomenological model then previously reported. With minimal loss in speech information and a robustness to noise, this paper concludes that WavIHC is suitable for use as a biologically accurate feature extractor.