The Stenographic Reports Department of the Austrian Parliament produces the stenographic reports of the parliament's plenary sittings. The crucial constraints, the department writes, are time and correctness, and the team had a long-standing dream of "a high-quality report in no time." Three coincidences led to the STT project: a stubborn MS Word crash during a plenary sitting prompted the team to replace the existing editorial toolchain; commercial STT providers had approached the parliamentary administration in 2018-2019 with no internal owner; and department member Julia Schöllauf was not only a trained parliamentary stenographer but, in the department's words, "very IT-savvy," giving the project both the linguistic and the technical owner it needed.
The department describes the project as "the first productive AI project in the Austrian parliament" and was designed bottom-up from the start. The Department, with its linguistic expertise, defined requirements and worked alongside the Parliament's IT department on infrastructure and interfaces; legal experts were embedded in the project team to head off compliance issues; and the starting question was specifically how an STT layer could enhance the existing editorial process, not how to replace it. A November 2023 conference in Vienna invited other German-speaking parliaments to compare notes, after which the team decided to pursue OpenAI's Whisper as the underlying engine, running experiments across multiple plenary speeches to confirm quality before committing.
The STT service in production since the beginning of 2026 is a multi-module pipeline. The audio stream of a sitting is automatically cut into sections aligned to the stenographers' workflow; those sections are sent to the STT service, whose core is three components: Faster Whisper (open-source, MIT-licensed, a speed-optimised version of OpenAI's Whisper using the same language model) produces the raw transcript; Pyannote (also open-source, MIT-licensed) detects speaker changes; and a third component does automated post-editing against the department's editorial standards, which uses no AI and is mostly based on searching and replacing patterns in the text with regular expressions. The corrected transcript is then sent to the existing editorial system, where the prior multi-stage human review was deliberately preserved unchanged: parliamentary stenographers remain in the hall writing down interjections, applause and procedural actions; typists act as a first-pass human control layer focused on obvious recognition errors and formal requirements; stenographers edit the typist-revised version; and a senior editor runs a final correction loop. No step of human review was removed.
The department says it is too early for a comprehensive evaluation, but early experience shows a significant acceleration of the process: "the former typists can now process about double the amount of text per hour," with reduced stress on long sitting days and, in the longer term, an expected reduction in the number of typists on duty per plenary day as retirements take their toll. The department also reports a quality improvement tied to the system producing consistent output even after many hours of work. It expects further improvement as the system is developed further.