Adding Fast Whisper
For speech-to-text
I have already added Android’s system voice and Android’s Whisper offline edge AI local model.
However, the AI computing power of mobile phones is too low, and transcribing even a short audio clip takes a long time.
Therefore, I need to deploy the service on a computer. Although it’s a P106-100 6GB headless mining card with an older chip architecture, it can still meet daily voice tasks.
I want to specifically test how many times faster the computer is compared to the phone.
Theoretically, a computer’s performance far exceeds that of a mobile phone, especially with a dedicated graphics card.
After installation and debugging, although there’s no specific data, it’s at least 2 times faster. I tested it on a local area network. If deployed to a public domain, it would also require time to upload audio, limited by network upload speed.
Ollama Translation
I need to translate articles into English, which requires a local LLM large model.
Similarly, I previously integrated the Gemini API and local LLaMA offline models.
But I still need to leverage the graphics card’s capabilities on the computer, truly pushing it to its limits.
Always adding a note: I rarely play games.
I chose the Qwen series large models: Qwen2.5 7B, Qwen3 8B, Qwen3.5 9B.
Currently, the 8B model performs decently, although it’s relatively slower.
Image Compression FFmpeg
I forced the app to select FFmpeg to handle images and videos.
Set images to retain dimensions and then compress while retaining file type.
Set videos to be convertible to GIF or WebP animated images.
Summary
In summary, I added computer-based open-source AI projects on top of the existing mobile edge offline AI and cloud API solutions.
To achieve faster note-taking:
Speech-to-text is at least 2 times faster. Text translation is 2 times faster. Image compression method optimized, speed is not a concern.
Also fine-tuned the button styles in the Markdown editor.
