Linguists are addressing the predicted the loss of many of the world's languages through an emerging discipline called Language Documentation, which focuses not on theory but on data, and how the data is acquired, represented, presented, and preserved. For most endangered languages, which are not written, much of this data is audio, and unlike many corpora it is likely to be local, particular, opportunistic, and uneven. New questions are raised, such as: what audio data counts as a record of a language that is likely to disappear? how can coverage and quality be measured? for what purposes and by whom will the data be used? For those of us documenting languages, there are four key audio-related issues: audio quality, its accompanying symbolic data, the usage of data for practical purposes such as language revitalisation, and the need for enhanced sensitivities and protocol in audio access and distribution. Language Documentation has benefited from the knowledge and experience of other disciplines, but perhaps it is now sufficiently experienced to offer some useful advice to others. This paper surveys these issues, and also describes the funding, teaching, archiving and publishing activities of the Endangered Languages Project at SOAS.