AssemblyAI has introduced "keyterm prompting" to improve the accuracy of its real-time speech-to-text models, particularly for specialized terms like names, jargon, and product names. This feature addresses the common issue where models correctly transcribe general speech but fail on high-value, rare words. Keyterm prompting works through two stages: real-time word-level boosting during inference and a post-turn phonetic matching process using Metaphone to correct misheard terms. This functionality is available on AssemblyAI's Universal-3 Pro and Universal-3.5 Pro Realtime models. AI
IMPACT Enhances the practical usability of real-time speech-to-text for specialized applications by improving accuracy on critical terms.
RANK_REASON The item describes a new feature for an existing speech-to-text service, not a core AI model release or research breakthrough.
- AssemblyAI
- Byrne-Donoghue
- Keyterm prompting
- Metaphone
- Metoprolol
- Universal-3.5 Pro Realtime
- Universal-3 Pro
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
