RAG for audio and meeting recordings: transcription first
Audio has to be transcribed before it can be retrieved at all, and how the transcript is segmented - by timestamp and speaker turn - matters more for retrieval quality than the last few points of word accuracy.
Transcription is the conversion
There is no way to embed sound directly for text retrieval. A speech recognition model turns the audio into text, and everything after that is ordinary RAG. Which means the transcript's quality is a hard ceiling on how good retrieval over that recording can be.
Timestamps turn an answer into a citation
A transcript chunked without timestamps can tell you a topic was discussed. One chunked per segment, with its start time carried in the metadata, can tell you it was discussed 41 minutes in - so a person can go and listen. That difference is most of the practical value of converting a meeting at all.
Speaker turns are natural chunk boundaries
A speaker change is usually a topic or stance change, which makes it a far better place to cut than a fixed word count. Chunking across turns produces passages that mix two people's positions and retrieve as though one person said both.
Watch for a fixture transcript
Some tools ship a placeholder speech engine so the pipeline can be tested without model weights. If your transcripts look suspiciously generic or identical across files, check whether a real model is actually loaded - a fixture transcript looks like a successful conversion right up until you search it.