Readable Meeting Transcripts That Keep the Detail

The challenge

Raw speech-to-text is hard to read. Sentences run together, punctuation is guesswork, and one misheard word can turn a fund name into nonsense. Passing that to an adviser, or to a model drafting a report, carries every error forward. Asking a language model to tidy it up brings its own risk: models like to summarise, and a summary of an advice conversation can drop the one caveat the client raised. Long meetings add a practical problem, because an hour of conversation is more text than can sensibly be cleaned in a single request.

What we built

A clean-up stage in FinPlan that sits between transcription and report drafting:

• A narrow brief for the model: GPT-4o works as a transcript editor at a low temperature of 0.1. It adds punctuation and capitalisation, forms paragraphs, fixes obvious recognition errors and keeps every speaker label. Its rules forbid summarising, omitting or adding content, and tell it to leave unclear passages as they are.
• Splitting at natural breaks: long transcripts are cut into sections of up to 32,000 characters. Each cut prefers a change of speaker or a paragraph break, then the end of a sentence, so sections rarely end mid-thought.
• Context on both sides: every section is sent with 2,000 characters from the sections before and after it, plus the list of identified speakers, so the model knows who is talking and what was just said.
• Parallel, then back in order: sections are cleaned at the same time and reassembled in their original order. A failed request retries up to three times, waiting longer each time.
• A check for lost text: a diagnostic compares the raw and cleaned transcripts for any record and flags it when text appears to have gone missing.

Technology

GPT-4o through OpenAI or Azure OpenAI, as described in Switching AI Providers Without Changing the Product, called from a Kotlin and Spring Boot service that uses Project Reactor for parallel requests. Section size, context length and retry settings are configuration values, so they can move with the model without a code change. We run this stage in production as part of An AI Platform That Turns Client Meetings Into Finished Reports.