Subjective effects of cell losses in voice over ATM

We present the results of formal subjective listening tests of voice transported via asynchronous transfer mode (ATM), with emphasis on cell losses. The μ-law voice samples are divided into pairs of ATM cells, one containing most significant bits (MSB) cells and the other containing least significant bits (LSB) cells. Based on isopreference testing, we conclude that: (1) isolated blocks of 15 or fewer consecutive LSB cell losses (176 ms or less) are indistinguishable from the original speech and (2) LSB cell loss rates up to at least 10% (10/sup -1/) are indistinguishable from the original speech.