RCRAG Converter

Evaluate local speech transcripts before you index them for RAG

Re-transcribing audio with an updated local speech engine is a corpus change, so each transcript should carry its engine version and pass a regression check against the previous run before its passages replace the old ones in a RAG index.

How do you check a speech engine update before re-indexing transcripts for RAG?

Treat the update as a corpus change. Record the engine, version and commit with each transcript, re-run a fixed audio set, diff the text and timestamps against the previous transcripts, and replace indexed passages only when every difference is explained. A speed claim is not part of that gate.

Transcription is the first conversion, and it has a version

The audio and meetings guide covers the first rule: audio has to become text before it can be retrieved, and segmenting by timestamp and speaker turn matters more than the last points of word accuracy. This article covers what happens next time. When the speech engine that made those transcripts is updated, every transcript it produces again is new input to your corpus, even when the audio has not changed.

So a transcript needs provenance. Store the engine name, its version, and where possible the source commit beside each transcript file, together with the model name and any decoding settings such as beam size. Without that record, a changed passage in your index cannot be traced to a changed engine, a changed recording, or a changed chunker.

A real example: a speech engine changed on its main branch

Mercury, the speech component of the open-source FFAI suite, is a useful case. In November 2026 its developers landed copy-reduction work on the FFAI master branch, after the published ffai-mercury 1.1.0 crate. On their own harness of 13 LibriSpeech clips, speech-recognition memset fell from 573.3 MB to 3.3 MB and attention-cache copies from 510.0 MB to 14.4 MB, according to the project's ledger. Those figures are repository-reported counters. The same ledger records the same-day timing run as inadmissible, so the work carries no speed claim, and it is not in a published release.

The useful part for RAG is how that ledger decided what to keep: counters had to fall and greedy and beam-search outputs had to stay byte-identical to the baseline on the harness. That is the shape of a regression gate you can copy. RAG Converter does not call Mercury or any speech engine on this path; the example is about evaluating transcripts you produce elsewhere.

Build a regression set before you re-transcribe

Pick a small, fixed set of recordings that represent your corpus: a clean read, a noisy meeting, a long file, a near-silent stretch. Public material such as LibriSpeech is a fair addition because anyone can rerun it. Keep the previous transcripts of that set as the baseline.

After an engine update, transcribe the set again and compare. Check, in order: whether the text is byte-identical; if not, which segments changed and by how many words; whether segment start and end times moved; and whether segment counts changed, because a changed boundary moves every chunk after it. Score word error against a reference where you have one; the NIST rich transcription evaluation pages describe how that kind of scoring has been organized. Only an explained difference should reach the index.

One identical run is not proof of determinism

Run the comparison more than once. The same Mercury ledger reports that its text-to-speech path, run 150 times per binary with a fixed seed, produced exactly two different waveforms: the alternate appeared in 5.3 percent of pre-campaign runs and 2.7 percent of post-campaign runs, a difference the ledger does not treat as significant. That finding concerns speech synthesis, not recognition, but the lesson transfers: a single matching run does not show that every run matches.

For transcripts, repeat the regression set at least twice per engine version and record whether the outputs agree with each other before you compare them with the baseline. If two runs of the same version disagree, fix or document that first. Otherwise a version comparison measures noise.

Hand the checked transcript to conversion as text

Once a transcript passes, save it as plain text or Markdown with one heading or paragraph per segment and the start time written into the text. The browser converter at ragconverter.com handles text and Markdown in its Preview conversion path, and the Markdown guide explains why headings make good chunk boundaries. Keep the provenance record next to the converted output so a later diff can name its cause.

Replace passages in your vector store through your own pipeline, following the incremental indexing guide, rather than rebuilding blind. RAG Converter's persisted corpus and query API is planned, not available, so the index you update is yours. The conversion stack comes from Remade with Rust, memory-safe libraries built for local use; that ecosystem is context for the converter, not a promise of faster transcription.

More on converting for RAG