Introduction
The crucial factors for creating parliamentary reports are probably the same all over the world: time and correctness. These factors are highly related, as quality requires time. So, every department in charge of parliamentary reports has one dream: a high-quality report in no time.
Background
The Stenographic Reports Department of the Austrian Parliament is no exception in this respect. In our endeavour to optimise our editorial process, however, three coincidences led to the project to implement an in-house speech-to-text (STT) system.
First, one late and stormy night during a plenary sitting, MS Word acted up again—and we decided to get rid of it. This led to a project that completely changed the editorial system of the stenographic reports—and, as a by-product, created an interface for STT software.
Secondly, in 2018 and 2019, when STT systems were not yet as effective as they are now, many providers approached the parliamentary administration. Someone had to take on the issue in their department. As the topic was not yet very popular at that time (no AI hype yet), it fell to the Stenographic Reports Department.
Thirdly, Julia Schöllauf in the Stenographic Reports Department is not only a parliamentary stenographer but—partly due to her family background—very IT-savvy. So, we suddenly had someone who is at home in both worlds: language and IT technology.
The latter proved to be extremely useful, as the STT project was designed bottom-up from the very beginning—our department, with its profound linguistic expertise and a clear picture of the requirements, being strongly involved in all steps. This significantly reduced the communication effort because a part of the system was developed directly in our department, while the IT department was responsible for more general issues such as infrastructure and interfaces.
As our project—the first productive AI project in the Austrian parliament—raised several legal questions to be solved, another measure to shorten the communication channels was to involve responsible legal experts in the project team to anticipate potential legal complications and address legal issues efficiently.
Another success factor was to start from the particular use case: how could such a system enhance our editorial process? Financial considerations were not the driving force behind the project but the question of how to cover more reports and transcripts with the same resources. Resulting economies due to faster processes were more of a side-effect than a key objective for us.
International exchange
To start with, we focused on knowledge sharing. We thought it was not necessary to reinvent the wheel and thus invited all German-speaking parliaments to a conference in Vienna in November 2023, bringing together parliamentary stenographers and IT experts to share their experience and views. One finding was that Whisper from Open AI provided promising results, so we decided to continue with this approach. Experiments with different speeches from plenary sittings confirmed the solid quality of the transcriptions. We decided to use Faster Whisper, which is a speed-optimised version of Open AI’s Whisper, using the same language model.
Having it in our hands
The promising results were only one of several factors that led to the decision to develop an in-house solution. Other factors were:
- Existing systems. As mentioned above, a new editorial system for the stenographic reports had been developed in-house, already providing an interface for a future STT service. Furthermore, there is an audio management system for the whole house which hosts the recordings of all sittings. An STT system by an external provider would have had to be integrated into these systems.
- Data security. A speech-to-text system for parliamentary reports must be suitable for all kinds of sittings, including those dealing with confidential and classified information. Therefore, it was essential that this information remained in our own infrastructure.
- Vendor lock-in. Many external providers approached us with their products. No matter how convincing the results, one problem remained: what if the company did not develop in the desired way—if it was, for example, sold to a big corporation where data security was no longer guaranteed? This was considered an unnecessary risk, given the existing in-house resources.
- Last, but not least, flexibility. Although the results provided by Whisper showed solid quality, it was soon clear that automated adaptations to the editorial standards and formal requirements of the stenographic reports were necessary in order to derive the maximum benefit from such a system. These adaptations were made directly in the Stenographic Reports Department, which would not have been possible with an external provider. Furthermore, we can constantly adapt the system to new requirements—such as new terminology and names—in an easy and timely manner.
Comprehensive system
Since the beginning of 2026, our STT system has been in productive operation. For all plenary sittings, stenographic reports are created in an adapted workflow involving the new system. In roughly one and a half years, the team of developers from the IT department and Julia Schöllauf from the Stenographic Reports Department developed a comprehensive system, consisting of several modules.
First, the audio recording is automatically cut from the audio stream of the sitting into sections according to the stenographers’ workflow. These sections are then sent to the STT service, which is the core of the system and consists of three components. Faster Whisper provides the raw transcript; Pyannote, an AI-based tool, detects speaker changes. Both are open-source and under the MIT licence, which enables usage with hardly any restrictions. The third component is automated post-editing according to editorial standards. It does not comprise any AI and is mostly based on searching and replacing certain patterns in the text (regular expressions). This optimised transcription is then sent to the editorial system and forms the basis of the multi-stage editorial process, which remained unchanged: the parliamentary stenographers are still in the hall, writing down everything notable, such as interjections, applause and activities. In order to guarantee the same level of human control, the typists are not replaced by the STT system but serve as a first control stage in the process, with a focus on obvious recognition errors and formal requirements. In the next step, the stenographers edit this first human-revised version, which is followed by another correction loop by a senior editor.
Conclusion
It is too early for a comprehensive and profound evaluation, but the first experiences in practice showed a significant acceleration of the process, which reduces stress on a long day of a plenary sitting. In the longer term, reduced costs are also expected, as the first stage of the editorial process requires significantly less time: the former typists can now process about double the amount of text per hour. As soon as the new workflow is well established, fewer of them will be on duty on a plenary day, which will be handy in the light of retirements. At the same time, quality improvements have been observed because the STT system provides consistent quality even after many hours of work. Further improvement is expected through further development of the system.
We are actively addressing the potential opportunities and challenges of the future. In this respect, we encourage all colleagues to take the initiative and shape their own destiny.
Julia Schöllauf is a parliamentary stenographer in the Austrian Parliament and the co-developer of the STT system.
Bettina Brixa is a head of the Stenographic Reports Department of the Austrian Parliament.
Dario Summer is a senior editor in the Austrian Parliament and the coordinator of the STT project.
Further reading
Faster Whisper. URL: https://github.com/SYSTRAN/faster-whisper.
Pyannote. URL: https://github.com/pyannote/pyannote-audio.
