Why Local AI? And What Is KB-Whisper?
Reading time: approx. 7 min
What You Will Learn
In this first moment we lay the foundation. You will understand why it is strategically important to start experimenting with local AI right now. We also introduce KB-Whisper, which is the Swedish AI model we will be using.
The Situation with Local AI
Using ChatGPT through the web is simple and powerful, that is true. But it naturally means you are sending your data to a company that controls both the model and the information. Running AI locally on your own computer flips this around and gives you several advantages:
- Full control and privacy: No data leaves your computer. You can work with sensitive information without worrying about third-party access.
- No costs beyond hardware: Open models are free to use. You do not pay per question or per user.
- Customization and specialization: You can fine-tune and retrain local models for specific purposes, for example for a particular school subject or terminology.
- Offline functionality: The model works without an internet connection once it is installed.
Sure, it requires more technical knowledge and a reasonably powerful computer, but development is moving fast. The skills you build today will be invaluable when local AI models become even easier and more powerful.
So What Is KB-Whisper?
The Royal Library has the mission to collect and preserve everything published in Sweden. Thanks to this unique position, they have been able to develop AI models that are deeply rooted in the Swedish language and culture.
KB-Whisper is such a model, a speech-to-text model that has been trained to convert spoken language into written text. The model is based on OpenAI's Whisper model but has undergone thorough specialized training on an enormous amount of Swedish material.
The Training Material That Makes the Difference
To become an expert in Swedish, KB-Whisper has been trained on over 50,000 hours of varied Swedish speech, including:
- TV broadcasts: Subtitled programs from SVT to capture a wide variation of spoken Swedish.
- Parliamentary debates: Speeches from members of the Swedish parliament.
- Dialects: Recordings from the Institute for Language and Folklore.
The result is a model that is superior to others in Swedish, with 47% fewer errors at the word level compared to the global original model. It is particularly good at recognizing and spelling Swedish place and personal names correctly.
How Can We Use This in the Classroom?
Even though the process we will go through is too technical for most students right now, it opens doors for you as an educator:
- Create teaching materials: Quickly transcribe an interesting YouTube video, a lecture, or an interview to create text-based material.
- Support for students: Create transcripts of oral presentations for students with reading and writing difficulties.
- Streamline administration: Transcribe meeting notes or other recordings.
KB-Whisper is free to download and use, but it lacks a graphical user interface. That is the reason this course exists.
Next Step
In the next moment it gets practical. We start by installing the tools needed to download audio from a YouTube video.