Transcription Quality
Measuring accuracy and formatting consistency in speech-to-text labels.
What this means for trainers
Follow the exact formatting convention specified in the guidelines rather than your own natural transcription habits — consistency across the dataset matters more than any single stylistic choice being "correct."
Transcription quality evaluates how accurately spoken audio has been converted to text, covering word accuracy, correct handling of disfluencies such as filler words and false starts, speaker attribution in multi-speaker audio, and consistent formatting conventions for punctuation, casing, and how numbers and abbreviations are written.
This work often overlaps with speaker diarization in multi-speaker recordings, where getting the words right is not enough if they are attributed to the wrong speaker, and with related tagging tasks like audio event labeling when non-speech sounds need to be marked alongside the transcript.
Formatting consistency matters more than it might seem. A dataset with inconsistent conventions for numbers, filler words, or punctuation introduces noise that downstream speech and language models have to learn around rather than learn from, even when every individual transcript is technically accurate on its own.
Put this into practice
Browse open data annotation and evaluation roles from Mercor, Micro1, Outlier, and more.
Browse AI training jobs