Single channel speech enhancement using adaptive filtering and best correlating noise identification

Autor: M S Athulya, Sathidevi Puthumangalathu Savithri, Vinayshankar Somalara Nataraj
Rok vydání: 2017
Předmět:
Zdroj: CCECE
DOI: 10.1109/ccece.2017.7946770
Popis: Speech enhancement using adaptive filtering methods are known to give good signal recovery from the noisy speech signal. Among these, Least Mean Square (LMS) and Recursive Least Squares (RLS) algorithms are more popular. These algorithms have a constraint that correlating noise should be given as the reference signal for denoising. Therefore in all the adaptive algorithms, two microphones are used, one for capturing noisy speech and the other for capturing noise signal alone. Always, capturing noise alone is difficult. To overcome the above constraint, we propose a novel method which identifies the best correlating part of the noise signal with respect to noise in noisy speech and this can be used as the reference for speech enhancement in adaptive algorithms. Once, best correlating noise is identified, Variable Step Size LMS (VSSLMS) and RLS algorithms are used for speech enhancement. Prior to this, noise classification is done to identify the type of noise present in the speech. Bark features and Support Vector Machine (SVM) are used for noise classification. SVM based noise classification combined with the identification of the best correlating part of noise for using as reference in adaptive algorithms for single channel speech enhancement is proposed in this work. Proposed system gives very good performance even in the case of speech mixed with non-stationary noise under very low signal-to-noise ratio (SNR) conditions.
Databáze: OpenAIRE