Performance of the Federal Standard 2.4 kbps MELP Vocoder Over ATM Networks

This paper evaluates the performance of the federal standard 2.4 kbps MELP vocoder when used over ATM networks that are subject to cell loss. It begins by addressing the sensitivity of MELP to random deletion of frames; it quantifies the extent to which the quality of the reconstructed speech is affected by two factors – the rate at which frames are lost and the burstiness of that loss process. Next, the ATM structure is introduced and techniques for ameliorating the effects of ATM cell loss when transporting MELP-compressed speech are evaluated. Interleaving MELP frames into ATM cells is shown to improve performance significantly. Finally, the effectiveness of erasure decoding using Reed-Solomon codes to recover lost cells is investigated. Throughout this paper, cell loss is simulated by an i.i.d. model and also a bursty (Gilbert) model. The quality of the reconstructed speech is assessed using both objective measures (i.e., frequency-weighted average spectral distortion and percentage of outliers) and subjective tests.