Customer story
About SpeechMind
SpeechMind writes formal meeting minutes for the public sector: town halls, councils, government sessions in Germany, Austria, and a handful of Swiss customers, who make up 70 to 80% of its customer base. The product is built around a simple but strict requirement: in a formal meeting record, the speaker's name has to sit in front of every statement.
SpeechMind didn't want a single bundled black box. Customers choose their own ASR model. They can choose their LLM; Mistral, for example, when a European option matters to them. SpeechMind runs its own normalization step rather than relying on any vendor's end-to-end orchestration. Diarization is the one piece they decided not to build, or buy pre-bundled: it runs on pyannoteAI, in parallel with transcription, then merges with it afterward.
Because they are focused on really formal meetings. It's important in this kind of meeting minutes that you can understand who said what.
Where attribution actually breaks
SpeechMind's hardest audio isn't a clean online call; it's the opposite. Most of it comes from in-person town halls and council chambers: a phone left in the middle of a room, 25 people talking, some of them interrupting each other.
Justus Feron
Co-Founder and CIO
Because the output is a formal record, not a quick summary, precision outranks speed. Overlapping speech that the system flags as uncertain goes to a human reviewer before minutes ship; everything runs in batch, after the meeting, and a next-week turnaround is fine with most customers.
Two separate features, two separate workflows
SpeechMind treats diarization and speaker identification as different products with different rules. Diarization (labeling who's speaking, even without knowing their name) runs by default. Speaker identification, which matches a voice to a known person, is opt-in and consent-based: a group can approve AI-generated minutes as one agenda item, but turning on identification requires every attendee to individually accept it.
Adoption of identification currently sits around 15% of customers, but it's rising, particularly after elections, when a new group of officials sets up their data preferences from scratch rather than inheriting an old decision.
Fitting pyannoteAI into an existing pipeline
SpeechMind first came to pyannote through the open-source model, via a transcription vendor that bundled basic diarization. As their customer base grew and audio got noisier and harder, so did the complaints, which is what pushed the move to pyannoteAI's commercial Precision model.
"We don't have a big labeled dataset to prove it with numbers. But at least the customers saw the difference, especially on hard cases with noisy data. And we saw fewer complaints once we switched. When customers stop complaining, that's basically our success metric."
Justus Feron
Co-Founder and CIO
SpeechMind runs diarization and transcription in parallel and merges the results, keeping the model swap self-contained rather than requiring changes elsewhere in the stack.
Why on-premises, not the API
The deployment model wasn't a technical afterthought. It decided which vendors SpeechMind could even consider. Its customers are politicians and public administrations, and they arrived with hard requirements: where data is processed, which model touches it, and whether any of it trains something else. Even a European vendor isn't automatically clean in a customer's eyes ; subprocessors get checked. "If we see with a provider that it's running on AWS, it's a European company, that's all nice. But at the end, they go look at your subprocessors and see they're in America. That's already a complaint," Justus says.
That's why SpeechMind runs pyannoteAI's diarization, voice identification, and speaker identification models fully on-premise. "Data security was the top priority, never leaving data in the US. That was the reason why we went on-premise."
How pyannote supports on-premise running team
Owning the deployment meant owning GPU provisioning, Docker, Python version upgrades, and dependency management. "Deploying the model and keeping it running was more effort than putting the results to use," Justus says. pyannote’s Support offset that: early calls with founders, responsive email support, and a same-day fix after a downtime.
What's next
SpeechMind is looking at using transcript context, not just audio, to catch and fix attribution mistakes in overlapping speech, and evaluating pyannoteAI's STT orchestration to simplify the handoff between transcription and diarization even further.
As European cloud infrastructure options mature, they're also open to moving parts of their pipeline to pyannoteAI's API, without giving up any of the sovereignty guarantees pyannoteAI already delivers on-premise today.
For SpeechMind, transcription tells you what was said. pyannoteAI is what tells you who said it, reliably enough to put a name on the record, and sovereign enough to run inside a government's own walls.
Discover more stories
