Theoretical accounts of both speech perception and of short term memory must consider the extent to which perceptual representations of speech sounds might survive in relatively unprocessed form. This paper describes a novel version of the serial recall task that can be used to explore this area of shared interest. In immediate recall of digit sequences, a recall advantage for the final item in spoken rather than written lists has been interpreted as evidence for a precategorical acoustic store (PAS). Three experiments are reported in which participants recalled lists of digits presented with synchronised speech and visual displays, with spoken items degraded by narrow band-pass filtering or by signal processing techniques that eliminate formant cues. Although simultaneous provision of top-down cues from the visual display creates an illusion of speech clarity, memory for terminal list items reflected the intelligibility of the speech signal rather than listeners’ perceptual experience. Recall of bimodally presented lists is apparently supported by information retrieved from an acoustic memory trace that is unmodified by top-down influences. This finding supports the PAS hypothesis, but is inconsistent with predictions generated by interactive processing models of speech perception.