Why structured Markdown is the right LLM input format for audio

A flat transcript is a wall of text. The LLM has to re-derive turn boundaries from prose ("Sarah replied that…"), guess at topic shifts, and invent citation anchors when asked for quotes. On a 60-minute meeting that re-derivation goes wrong often enough to make answers unreliable.

Markdown with ## Speaker [HH:MM:SS] headings gives the model three things at once: who is talking (heading text), when they talked (timestamp), and where the turn ends (next heading). Every modern LLM — GPT, Claude, Gemini, Llama, Mistral — was trained on enough Markdown to treat heading boundaries as semantic. Plain text gets none of this for free.

Semantic chunking, finally working

RAG over audio used to require custom diarisation pipelines and per-speaker chunking heuristics. With structured Markdown output, chunking is one line: split on ## and each chunk is a coherent speaker turn. Embeddings then cluster on what was said rather than averaging across speakers, and retrieval surfaces the actual relevant exchange instead of scattered fragments.

Model-specific guides

• ChatGPT — speaker attribution and meeting analysis
• Claude — Projects-as-meeting-archive patterns
• Gemini — controllable input for the 1M-token window
• RAG — podcast and meeting knowledge bases
• LangChain and LlamaIndex — code-level integration

For PDF and URL sources, see PDF to Markdown for LLMs and URL to Markdown for LLMs — same principles, different input formats.