Smart systems and artificial intelligence technology are becoming increasingly popular and are continuously finding more applications in real-life situations. Many systems require human-computer interaction and the natural language interface. One of the major issues in speech recognition systems is their performance in real world (noisy) environment. Over the past decades many techniques for noise reduction were developed. Motivated by human auditory processing, and it is well known that humans are remarkably good at detecting speech in the background noise, we propose a noise reduction technique based on a biophysical cochlear model. Using a model of signal reconstruction from the cochlear output, we observed an improvement in the quality of noisy speech and a significant increase in speech recognition performance.