Prepare a request for transcription of this recording. Return JSON {"glossary":["..."],"speakers":"...","output_needs":"..."}: a glossary of names/terms to tell the model how to mark speakers, and what is needed in the output (time codes, marking uncertain places, etc.).
Transcription - deciphering a long recording into text. Under the hood, the model not only recognizes words, but also tries to separate speakers (diarization) and arrange tenses. Accuracy falls on three things: unfamiliar names and terms, overlapping speech (interrupting each other), and strong accents or noise. Hence the leverage. The first and most underrated one is to give the model a glossary in advance: a list of participant names, titles, abbreviations. Words “presented” in advance are recognized noticeably more accurately than words guessed from scratch. Second, ask for markings by speakers and by time, so that you can quickly find the right moment. Pro trick: if the entry is important, ask to mark uncertain places with brackets or a sign - this way you can immediately see what to double-check, instead of trusting the smooth, but possibly erroneous text. Insider: 80% of the quality of transcription is determined by the quality of the input sound - one good microphone in the center of the table is better than any post-processing. A typical beginner mistake is to post a two-hour recording without a single name or context, and then clean out dozens of errors manually.
Unlock access to submit solutions for instant AI review.