Prompt engineering for Universal-3 Pro: A practical guide

The guide on prompt engineering for Universal-3 Pro, authored by Martin Schweiger, offers practical insights into using prompts to enhance transcription behavior and accuracy in specific domains without the need for custom model training. Universal-3 Pro merges traditional automatic speech recognition with the adaptability of instruction-following models, allowing users to control transcription through natural language prompts. The guide details how prompts can effectively manage six key areas: context and domain adaptation, verbatim transcription, spelling and terminology accuracy, multi-language handling with code-switching, speaker information, and audio event tagging. By providing clear, concise prompts and concrete examples, users can improve transcription outcomes, including reducing Word Error Rate and increasing domain-specific accuracy. The guide emphasizes starting with a base prompt and incrementally adding specific capabilities, recommending a temperature setting of 0.0 for optimal transcription accuracy while testing each capability to ensure the desired behavior is achieved.