How Much Time Does AI-Assisted Transcription Save? A Retrospective Telemetry Study of 3,999 Production Jobs
Year: 2026
Type: preprint
Abstract: Vendors of AI dictation and transcription services commonly advertise time savings of up to 80%, but the claim is rarely backed by a measurement that an outside reader can check. We report a retrospective analysis of 3,999 completed transcription jobs (785.9 h of audio, 5.15 million words, 138 users) recorded by a commercial German-first transcription service over 180 days. The median time from speech-recognition worker start to completion of LLM-based post-processing was 70 s (p90 352 s), and the median effective speech rate was 113 words per minute. We define two reference processes—archival transcription of an existing recording and live dictation—and derive time savings against manual typing plus proofreading under explicit, varied assumptions. In the base case (40 wpm typing, 200 wpm proofreading), the median saving across all jobs was 69%. For recordings of at least 20 minutes, the median was 80% (IQR 76–82), and 50.9% of jobs reached at least 80%. For live dictation, the median was 52% even for long recordings. The 80% figure is therefore supportable only for long-form archival transcription, and it depends on proofreading speed: the median long-form job requires proofreading at 188 wpm or faster. We report the complete sensitivity grid, including settings under which the claim fails, and propose minimum reporting requirements for public time-savings claims.
Author: Stein, Steffen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235237023
