Real example · Voice, avatars & transcription
Chose a transcription engine by testing ten providers
What they did
Fyxer, which makes an AI email assistant and meeting notetaker, benchmarked ten speech-to-text providers on its own meeting recordings, ran a live A/B test, and then made ElevenLabs Scribe its only transcription provider.
How it went
Fyxer reports a 20% lower speaker-attributed word error rate than its control and trial-to-paid conversion of 20.3% versus 17.6% across 6,000+ organizations, a 15% relative lift.
Worth knowing
Published by ElevenLabs, so it is vendor-reported. The audio is private, no absolute error rates or test length are given, and the old provider is not named.
Try it yourself with ElevenLabs
Replace anything in [brackets] with your details.
Help me design a fair test to compare [number] speech-to-text services for [use case]. Suggest a sample of [50] real recordings that covers accents, background noise and several speakers, how to create reference transcripts, the metrics to compute (word error rate, speaker labels, names and numbers), and a table to record results.Read the original
Fyxer uses ElevenLabs Speech to Text to power meeting notetaker ↗
elevenlabs.io · · Vendor case study
Attributed to the source linked below; not our own test.
Share Share on X Share on WhatsApp
