AssemblyAI has identified a critical flaw in keyterm prompting for transcription models, where adding specific terms can inadvertently degrade accuracy. Instead of acting as a simple dictionary, keyterm lists function as a bias, influencing the model's output even when it correctly transcribes audio. This can lead to incorrect transcriptions, particularly with proper nouns and product names that are acoustically similar to provided keyterms, as demonstrated when a correct transcription of 'Kelly Byrne Donahue' was altered by the presence of 'Kelly Byrne-Donoghue' in the keyterm list. AI
IMPACT Highlights a potential pitfall in using keyterm prompting for transcription services, advising users to be cautious of accuracy degradation.
RANK_REASON Blog post detailing a specific failure mode of a common AI tooling technique.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
