Real-time Swedish speech recognition – with audio processed in Sweden
With Klang AI's model and Berget AI's API, I built a prototype for real-time Swedish transcription in one evening. Here is why tech leaders should take a look.
On 23 September, Klang AI released Pianissimo, a Swedish speech recognition model with open model weights. The same day, Berget AI launched an API for real-time transcription built on the model. Together they offer real-time Swedish speech recognition where, according to Berget, the audio is processed in Sweden.
I wanted to see what could be built with it and built a prototype the same evening. What surprised me most was how quick it was to set up and get working. The prototype transcribes a Swedish conversation as it happens. Using a separate speaker model, I also added speaker labels, like Speaker 1 and Speaker 2.
The voices in the demo are generated with ElevenLabs. The demo shows the functionality; quality in real conversations needs to be evaluated separately.
The value is in what happens during the conversation
With real-time transcription, the text is available while the conversation is still going on. That lets an AI agent use it to look up information, suggest next steps and prepare tasks right away.
My prototype shows transcription and speaker separation. The next step is connecting the text to an agent and the business's systems.
In customer service, the agent can find relevant answers in the knowledge base while the customer describes the problem, and prepare the documentation during the call so there is less after-call work.
In meetings, the agent can suggest a task when someone says "Erik will put together a quote by Friday", look up details in the CRM, and prepare an update to the deal for the participants to confirm.
If you only need a summary afterwards, transcribing after the fact may be enough. Real time adds the most when the information can be useful before the conversation is over.
Buy or build?
If you only need to record and transcribe meetings, there are good off-the-shelf tools. Buy when they meet your requirements.
When transcription is part of an automated process, start by looking at which integrations are available. If they are not enough, it can be worth building the connection to your own systems and workflows. Weigh the benefit against the cost of developing, running and maintaining the solution.
Choose Swedish providers when you can
Meetings and customer calls can contain highly sensitive information. I think Swedish companies should consider Swedish AI providers far more often than they do today. With Pianissimo and Berget there is now an option where, according to Berget, the audio is processed in Sweden and neither audio nor transcripts are stored on their side. The model can also be run in your own environment.
That gives you more options, but the whole data flow needs to hold together. If the text is passed on to an AI agent or another system, you also need to know where it is processed and what is stored there.
Next step
Raise it internally: where would it make a difference if AI could act while the conversation is still going on? Then let someone on the team test a well-defined flow. My first prototype took an evening. A test like that can show whether the idea is worth taking further.
Want to see the prototype first? Get in touch and I will show it to you.
Frequently asked questions
- What is Berget AI's real-time transcription?
- An API that turns spoken Swedish into text while you talk, built on Klang AI's Pianissimo model, which has open model weights. Berget states that the audio is processed in Sweden and that neither audio nor transcripts are stored on their side.
- What is the value of real-time transcription?
- An AI agent can use the text while the conversation is still going on, for example suggesting answers and preparing documentation in customer service, or suggesting tasks and looking up information during a meeting.
- When is it worth building your own transcription solution?
- When off-the-shelf tools do not meet your requirements for data handling or integrations. If transcription is part of an automated process, it can be worth building the connection to your own systems, provided you can also maintain the solution.