Clean text in, better answers out

Garbage in, garbage compiled

Speech recognition mishears proper nouns, technical terms and unfamiliar names, and it leaves in the fillers and false starts of real speech. On its own that is a readability annoyance.

In a knowledge base it is worse. Compile a wiki page from a transcript that spells a name three different wrong ways and the page cannot link that name to anything - the connection that was the whole point is lost, and a citation that traces back to a garbled line is no citation at all.

Karpathy cleans his own transcripts

When Andrej Karpathy turned one of his talks into text, he did not stop at a raw transcript. He had an LLM read it and produce a cleaned version - correcting transcription mistakes, removing fill words - and published it for two audiences at once: people who would rather read, and LLMs, so that his content is legible to them. [1]

The cleanup is not cosmetic. It is what makes the text usable as a source instead of a rough approximation of one.

What Vocateca does before the text lands

Vocateca does that correction at the transcription step, not as a manual pass afterwards. A glossary of your own terms and names corrects mishearings by phonetic match, so a recurring name lands spelled the same way every time. A prompt biases the model toward the vocabulary you care about, and you can pick the engine that handles your audio best.

The transcript that reaches your archive is closer to clean out of the box - the same cleanup Karpathy did by hand, moved to where the text is made.

Clean text, better retrieval

Correct names mean correct links: a consistent spelling is what lets a concept connect across a podcast, an article and a talk. Correct text means correct citations: a sentence traces back to a source that actually says it.

And it means better answers. Whether a model reads the transcript directly or queries a network compiled from it, the ceiling on the answer is the quality of the text. Clean it once, at ingest, and everything above it improves.

Andrej Karpathy is not affiliated with Vocateca and does not endorse it. The reference above is cited for commentary.

Sources

1. [1]Andrej Karpathy, 'Sequoia Ascent 2026 summary,' karpathy.bearblog.dev, 30 April 2026. https://karpathy.bearblog.dev/sequoia-ascent-2026/ Accessed 11 July 2026.