Measured performance of block-floating-point hardware FFT processor for real-time speech transform coding

In order to explore experimentally the design of speech transform coding schemes it is necessary to implement forward and inverse FFT processors of sufficient accuracy that their contribution to the total distortion, over and above that of the source digitisation, is negligible. In this way we ensure that examination of rate-distortion coding tradeoffs is not obscured by additional distortion in the analyser/synthesiser subsystem. Accordingly, forward and inverse FFT processors have been designed and constructed incorporating an overflow correction scheme previously only implemented in software, which maximises their dynamic range. Objective measurements indicate a 49dB dynamic range for the analyser/synthesiser combination, and subjective tests confirm the negligibility of the concomitant distortion.