Audio Transcription Cost Calculator

Transcription is cheap per minute and expensive per pipeline, and the split above is the point of this page. The audio-to-text charge is the one people budget for; the tokens it produces are the one they forget. An hour of speech runs to roughly ten thousand tokens, and every pass over that text is billed at LLM rates — summarise every call, extract the action items, then answer questions about the archive, and the reading costs more than the listening. The defaults above model a single pass. Multiply the tokens-per-minute field by the number of passes you actually make.

Two structural traps. The first is per-request minimums: if your provider rounds every call up to a floor and your workload is thousands of short voice notes, you are paying for silence you never sent. The rounding row shows the uplift, though it uses your average file length and therefore flatters a workload with a wide spread. The second is re-transcription. Audio is immutable, so a transcript should be stored and never regenerated — but pipelines that treat the model as a function frequently re-run the whole archive when someone changes a prompt downstream.

What this does not price: accuracy. A cheaper model with a higher word error rate costs less per minute and more per useful answer, because the errors propagate into every summary and every search built on top. If you are choosing between transcription options, the honest comparison is cost per corrected transcript, and correction is either a second model pass or a human — both of which belong in the fields above as extra downstream tokens or as a separate line entirely.