Time delay estimation for objective speech quality

The effectiveness of various time alignment algorithms associated with speech is investigated. The delay estimation techniques considered here have been tested using several sets of speech data with varying types of distortion. Simple cross-correlation works best for the case of added noise, simulating a waveform coder for which it is optimal. However, for nonwaveform distortion such as T-ref simulating a low bit-rate codec, it is much less reliable. The envelope correlation seems to be consistently good for all the cases whereas the average zero crossing method seems to function more reliably for T-ref condition than the added noise case. Although, the cepstral distance method is reasonably accurate, it is computationally very intensive. Another drawback of this method is that it is not inherently a self-starting method, that is, to reduce the amount of computation with the cepstral method an approximation for the delay has to be made using one of the other methods prior to using this method. It could probably function best as a verification method. The statistical methods are less accurate than expected.<>