Real time and embedded implementation of hybrid algorithm for speech enhancement

The speech enhancement techniques are required to improve the speech signal quality without causing any offshoot in many applications. Recently the growing use of cellular and mobile phones, hands free systems, VoIP phones, voice messaging service, call service centers etc. require efficient real time speech enhancement and detection strategies to make them superior over conventional speech communication systems. The speech enhancement algorithms are required to deal with additive noise and convolutive distortion that occur in any wireless communication system. Also the single channel (one microphone) signal is available in real environments. The short time spectral amplitude or attenuation (STSA) based algorithm particularly minimum mean square error-log spectral amplitude (MMSE-LSA) is used in practice for background noise suppression. However, it is generating artifact called musical noise and has poor performance in low SNR (0–10dB), reverberant and nonstationary noise backgrounds. Hence an algorithm which has a dual capacity to enhance speech corrupted by additive noise and degraded by reverberation is desired. The relative spectral amplitude (RASTA) algorithm is used as a preprocessor to improve performance of speech recognition systems in reverberating and noisy environment. It can be interactively combined with STSA approach. This paper evaluates such hybrid algorithm in terms of objective and subjective measures. The real time implementation is done on PC using SIMULINK from Mathworks and embedded implementation is done on TMS320C6713 DSP from Texas Instruments using DSP Starter kit- DSK 6713 from Spectrum Digital Incorporation. The profile report of both the implementations are generated and compared.