Reference free speech quality estimation for diverse data condition

Abstract

The performance of any speech based systems depends on the quality of input speech signals. The signal to noise ratio (SNR) is considered to be a measure of the quality of speech signal. This paper reports some analyses (based on experimental evaluations) that are focused on calculating the quality of the speech signal so as to improve the overall accuracy of the system. As compared to existing methods that are based on voice activity detection (VAD), the proposed method is based on glottal activity detection (GAD) to detect the speech and non-speech regions from the input speech signals. Literature reveals that the GAD provides better results than VAD under noisy data condition for the above mentioned task. The proposed method uses a filter with specified cut-off frequency to separate the noise components that are present in the speech activity region. After that, we calculate the SNR as the ratio of the total energy in the speech activity region to that of the non-speech regions. The comparative analyses with two state-of-the-art techniques, viz, National Institute of Standards and Technology (NIST) SNR tool and Waveform Amplitude Distribution Analysis (WADA) SNR algorithm suggest that under different data conditions the proposed method outperforms the existing ones. Since the proposed model depends on the signal processing techniques and does not require any classifier or model to be developed, therefore it is found to be computationally efficient as compared to other two methods.