Auditory working memory (AWM) is critical for the temporary maintenance of sensory information, particularly in real-world communication contexts. Despite extensive research on visual working memory, the encoding of auditory features in working memory and their neural representation have remained less explored. As previous work has suggested that sounds are stored in working memory as objects including both task-relevant and -irrelevant features, we investigated whether we can decode both types of features from magnetoencephalography (MEG) signals. We conducted an experiment using a delayed match-to-sample task. Stimuli were spoken syllables that differed both in perceived spatial location and voice identity (i.e., specific combinations of timbre and pitch). Participants were instructed to memorize voice identity (the relevant feature), while location (the irrelevant feature) could be ignored. The task required a match non-match decision about a probe sound whose irrelevant feature value could either be identical to or differ from the memorized stimulus. Behavioral results demonstrated that voice identity matches were detected more accurately and faster when the irrelevant location matched the probe as well, whereas for voice identity non-match trials, there was no effect of the irrelevant feature. Multi-variate pattern analyses revealed that both task-relevant and -irrelevant features were decodable during the encoding phase. While the task-relevant voice identity yielded higher classification performance compared to location, location information was decodable earlier than voice identity. In conclusion, both our behavioral and MEG data suggest that location information is encoded alongside voice identity, and task relevance modulates the decodability of auditory features.