|
|
|
|
|
|
by jwr
23 hours ago
|
|
|
Incidentally, I've been testing a lot of dictation apps and models. I am on macOS Sequoia and intend to skip Tahoe altogether, so newer Apple models were not an option. I also dictate in multiple languages. I mostly use the MacWhisper app until now, but I got frustrated by its slow development and lack of focus and dictation. So I tried the free TypeWhisper app. It works very well, but then I tested multiple models and my conclusion was that one cannot beat the Whisper Large v3 model for accuracy. I tried Qwen ASR (very good, but only for English), Voxtral (spoke Polish to it and it wrote Russian, BIG oopsie), Parakeet TDT 3 (very fast, but sometimes inaccurate). I got frustrated by the speed of the Whisper Large model. So I spent an hour with AI benchmarking, testing and developing a new plugin. Turns out that on a modern Mac with a good GPU you can run the Whisper Large model at roughly the same speed as the tiny Parakeet model. 2-3x faster than what other apps ship. It's just that nobody bothered. So, I am now happily dictating into TypeWhisper and then feeding the dictated text into a local gemma-4-26b-a4b-qat LLM for corrections. Everything happens instantly. |
|
|
This is exactly why I made TypeWhisper’s engine and plugin layer extensible. The best setup depends heavily on language, hardware, and workload. Your result that Large v3 can run close to Parakeet speeds on a modern Mac is especially interesting, as is your fully local Whisper plus Gemma correction workflow.
If you’re comfortable sharing the plugin or benchmark setup, I’d be very interested in taking a closer look and seeing whether it could benefit more users. Either way, thank you for putting so much time into this and contributing your findings back to the project.