Graphics Card
My graphics card is a p100-100 6GB.
It’s a super old architecture chip graphics card that doesn’t support new AI features.
However, it still supports simple text, images, speech, local large models, and so on.
Speech-to-Text
This is also what I use the most; I’ve set up fast whisper in the app.
The recognition speed is very fast, and the quality is relatively good.
If it’s just for recording daily notes, then speech-to-text is super convenient, boosting efficiency by more than ten times.
I’ve changed my old way of writing blog posts on a computer to the current way of lying in bed and writing them.
It’s much more convenient.
I consulted AI, and it said that speech-to-text, real-time meeting minutes, and audio-to-subtitles are relatively common application scenarios.
Translation
Speech-to-text is suitable for specific groups of people.
But translation is essential for everyone.
Currently, for translation, I use local large models, such as Ollama, combined with Qwen2.5 7B. The translation effect and speed are relatively good.
This is my backup plan for when my Gemini quota runs out.
No Graphics Card
If you don’t have a graphics card,
you can also consider whisper.cpp and ollama.cpp. They only require a CPU and can be deployed and used on embedded devices like Android and development boards.
I haven’t tested their performance yet, but they will also serve as a backup plan later.
Summary
Currently, it’s just speech-to-text and translation; other things will be discussed later.