Subspace and Frequency Domain Speech Enhancement Techniques

Abstract

Speech enhancement or noise reduction is used as front end processing for speech recognition application. Speech enhancement applications include mobile phones, hand free phones, hearing aids, personal assistants, home automation, robots and so on. Also the hearing aid plays important role for hearing impaired listeners for comfort listening. To understand the speech enhancement algorithms it is important to analyze the output/performance by varying the parameters involved in the technique / algorithm. The main objective of paper is to compare different frequency domain approaches and time domain approaches available for speech enhancement. Karhunen-Loeve transform (KLT) and the MMSE estimators for speech enhancement is discussed. It is observed that considering perceptually motivated techniques shows improved performance and thus results are compared for basic approach and perceptual motivated approaches. This work discusses the theory related to speech enhancement and gives the guidance on how to proceed for implementation of speech enhancement algorithms using MATLAB. The real time application of mathematical operations like Fourier transform, Averaging, variance, Minimum Mean Square and windowing is discussed. Sub space algorithms for speech enhancement are discussed and the performance is compared with frequency domain approaches. Simulations are performed using MATLAB and the performance is compared using objective performance measures Signal to Noise Ratio (SNR), Segmental SNR and PESQ.