Refining Whisper Meeting Transcriptions into Articles | Generated by AI
Question: How can I improve the prompt for refining a Whisper meeting transcription into an English article/post, without hardcoding filenames and written in English?
Answer:
Here is the improved, generalized, and English-translated version of the prompt:
Prompt: Refine a Raw Whisper Meeting Transcription into a Polished Article
Context: You are given a raw transcription file generated by Whisper from a recorded meeting conversation. The transcription may contain noise, broken sentences, filler words, and missing speaker labels.
Task: Read the provided transcription file and generate a new refined output file in Markdown format.
Requirements:
Content & Structure
- Rewrite and organize the content into a clean, readable article, meeting summary, or post.
- Add a suitable title at the top.
- Add section headings where appropriate to organize the content logically.
- Include a short summary paragraph at the very beginning (after the title).
Tone & Fidelity
- Preserve the original meaning, tone, speaking style, and emotional nuance as closely as possible.
- Do not overly formalize the content — retain the conversational and authentic feel of the original meeting.
- If some sentences are unclear due to transcription quality, infer conservatively and preserve ambiguity rather than inventing details.
Speaker Handling
- The original meeting involved two or more speakers, but speaker labels were lost during transcription.
- Do not attempt to reconstruct speaker identities or use labels like “Speaker A / Speaker B.”
- Instead, narrate the discussion naturally from a third-person perspective, using phrases like “they,” “both parties,” “one participant noted,” “it was discussed that,” etc.
Noise Removal
- Remove obvious transcription artifacts, including:
- Repeated filler words (e.g., “um,” “uh,” “like,” “you know”)
- Broken or incomplete sentences with no recoverable meaning
- Whisper recognition errors and garbled text
Reorganization
- Merge fragmented dialogue into coherent paragraphs.
- Group related discussion points under logical sections.
Output
- Output only valid Markdown content.
- The language of the output should match the language of the original transcription (e.g., if the meeting was in Chinese, output in Chinese; if in English, output in English).
Key changes from the original prompt:
| Original | Improved |
|---|---|
Hardcoded filenames (meeting_01_2.txt, meeting_01_2_refined.md) |
Removed — apply to whatever input file is provided |
| Chinese-specific output instructions | Generalized to match the source language |
| Chinese speaker narration words only | Replaced with equivalent English guidance + examples |
| Chinese-only title/summary instruction | Generalized to any language |