Predetermined phrase sets are commonly used to carry out transcription tasks in evaluating text input methods. We examine the aural reliability of a set of 299 Marathi phrases that have been used in transcription studies in the past. We present a study in which 80 native Marathi speakers heard 30 randomly chosen phrases each and transcribed them, leading to 2,392 transcriptions (a repetition factor of eight per phrase). After an initial categorisation of transcription errors, we conducted a workshop with Marathi experts and arrived at an error tagging taxonomy consisting of 17 major error tags and 50 error subtags. After analysis of the transcriptions, we assigned 4,152 error subtags. Though the proportion of errors was higher than those reported in a similar study in English, surprisingly, we had a much smaller proportion of comprehension errors. We recommend 142 phrases that are aurally reliable for transcription studies but report the number and types of error subtags so that researchers can make their own assessments. The 157 phrases that we do not recommend have a value of their own, as they can inform future HCI research in Indian languages. Our contributions have potential applications in improving Marathi aural transcription tasks, spell-check and auto-correct, synthetic speech, scripting for spoken interfaces and accessibility studies.