Performance analysis of SS based speech enhancement algorithms for ASR with Non-stationary Noisy Database-NOIZEUS

In Human computer interface the correct translation by computer and exact perception of human depends on the quality of speech input to the machine. Hence speech enhancement technique is very essential in any HCI technique. In this paper we present an algorithm for speech enhancement on the non-stationary noisy database (NOIZEUES). The speech enhancement technique, Spectral Subtraction technique falls under the category of methods based on short time spectral amplitude (or power) estimate. This algorithm is found to be very simple and computationally efficient. Various modifications are made in the spectral subtraction technique and therefore spectral over subtraction, improved spectral over subtraction, iterative spectral over subtraction and multiband spectral subtraction techniques are derived. In this paper basic performance of spectral subtraction, spectral over subtraction and improved spectral over subtraction technique and SS with cross spectral terms with non-stationary noisy database is discussed. It is observed that algorithm works effectively to reduce additive noise present in the noisy speech signal. Limitations of spectral subtraction such as dependence on VAD accuracy and musical noise are described in the proposed paper. All listed algorithms are executed on available database experimentation results of speech and enhanced speech are provided in result section of this paper. Also numerical results are displayed graphically.