SAGARIS on how accurate AI note takers really are

Because they transcribed different audio. Word Error Rate is a property of a recording and the engine that processed it, so room noise, accents, crosstalk and speaker count move it far more than the product does. Two published reviews of one tool can disagree by about twenty points without either being wrong.

It depends entirely on which words the missing 5% were. WER weights every error the same, so dropping a price, a date and a name scores identically to mishearing three filler words. Ask what was measured and on what audio before treating a percentage as quality.

No, and this is the most common misreading. Speaker attribution is diarization, scored separately, and it depends on whether the engine that resolved for that recording performs it. A perfectly accurate transcript can still be a flat wall of text with no idea who said what.

Which engine transcribes my audio, was diarization on for this recording, what audio was your number measured on, and what happens when transcription fails. The last one matters most: a tool that always returns text may simply be filling gaps rather than admitting a miss. If the tool on your shortlist is Gong, or the conversation intelligence inside Clari and Salesloft, our Gong alternatives and Clari alternatives pages compare each with SAGARIS, with the date each of their facts was read.

We are not going to print one here, because the argument above applies to us too. We keep a Word Error Rate harness that scores transcripts against a committed set of reference texts. When a transcription provider is configured it checks real output against a floor; otherwise it skips. It is a measurement harness, not a live gate on your calls. In the in-person recorder, when transcription cannot run we return no transcript rather than generated text, and tests hold us to that. So no: we do not publish a SAGARIS accuracy percentage.