Attributing utterances to speakers in a good quality recording of clear speech might seem to be a simple and straightforward task, yet transcribers have been shown to “regularly and obliviously get it wrong” (Love 2020: 156). This issue has received very limited attention, and our research focuses on the issue from a forensic linguistic perspective where attributing an utterance to a speaker in a forensic case could effectively be accusing them of having committed a crime. Literature on speaker attribution/recognitionnever shows 100% accuracy even in very good acoustic conditions, with marked drops in performance when features are masked (as they often are in forensic situations). In this preliminary study, five phonetically trained listeners transcribed an audio file of six people engaged ina panel discussion from television (we focus on audio only). The speech of two female speakers is correctly attributed in almost all cases, so we pay attention to the acoustic features of four male voices. Listeners werealmost never correct in their judgements about the exact number of speakers in the recording and made multiple attribution errors. We show that this is because of the fundamental nature of the speech acoustics; within-speaker variation is larger than between speaker variation for F0, intonational properties, and speech rhythm. Love, R. (2020). Overcoming challenges in corpus construction: The spoken British National Corpus 2020. Routledge Advances in Corpus Linguistics. New York: Routledge.