Swiss German Speech-to-Text

A local speech-to-text project for Swiss German and German meetings. It transcribes spoken discussions, prepares summaries and lets users ask questions over sanitized transcript examples, while keeping confidential audio and text inside controlled infrastructure.

Built the transcription, summarization and transcript Q&A pipeline on controlled infrastructure.

Many important meetings contain decisions, prices, risks and commitments, but the person leading the discussion often cannot take complete notes at the same time. Standard transcription also struggles with Swiss German dialects.

I built a local transcription pipeline focused on Swiss German and German meeting audio, with summarization and a transcript Q&A layer so users can turn recordings into usable follow-up material.

The project works best for online meetings with clear audio and is still evolving for more difficult on-site setups. It already shows how spoken knowledge can become searchable, summarized and reusable.

Make spoken meeting knowledge usable without sending confidential audio to cloud services. The system should transcribe meetings, summarize them into department-specific formats and allow users to ask questions over the transcript afterwards.

The value is not only a transcript. It is the ability to recover decisions, action items, risks and follow-ups after the meeting. The public version uses generic examples only; real recordings, transcripts and outputs remain confidential.

Swiss German is the real challenge

The value of this project depends on Swiss German. A transcription system that works only for clean English audio would not help much inside a Swiss company where many meetings happen in dialect.

Basel, Zurich and other regions sound different, and real meetings add mixed languages, interruptions, distance to the microphone and domain-specific words. Online meetings with clear audio already work much better than on-site rooms with one shared microphone. The remaining challenge is partly AI and partly hardware.

Two models, strictly separated

One transcript, many documents

Ask the meeting, afterwards

Working, but still improving

The project is still in progress. Current results are strongest for online meetings with clear audio, because every speaker is captured cleanly. On-site meetings are harder when one shared microphone has to capture people at different distances and volumes.

The next goal is to make the output more adaptable for departments. Different teams need different summaries, templates and question styles, so the prompting and reporting layer should be configurable instead of one fixed format for everyone.

What it delivers