Voice typing has become part of how I use my M2 Mac. With VoiceInk, I press a shortcut, speak, and release it. The text appears so quickly that it almost feels instant.
The transcription is decent, although it does not always handle my Indian accent as accurately as I would like.
That left a gap on my Ubuntu laptop. VoiceInk is a native macOS application and is not available for Linux. I wanted the same push-to-talk workflow, local transcription, and good Wayland support.
Vocalinux turned out to be a good fit. It is much slower on my older Linux hardware, but Whisper Small often produces better text for me than Parakeet does on the Mac.
VoiceInk on the M2 Mac
I use the version of VoiceInk that I build from source. That is simply the way I chose to use it.
My local transcription model is Parakeet. On the M2, there is barely any pause between releasing the push-to-talk key and seeing the result. That speed makes voice typing feel like a normal input method rather than a separate process.
The accuracy is good enough for regular use, but I still need to correct words it misunderstands. This happens often enough with my accent that I notice it, even though I like the overall experience.
Vocalinux on Ubuntu
My Ubuntu laptop is considerably older. It has an Intel Core i5-8250U, 16 GB of RAM, and an NVIDIA GeForce MX150 with 2 GB of VRAM.
Wayland support was an important part of my choice. Vocalinux works well with my Wayland desktop and gives me the same basic workflow: hold the push-to-talk key, speak into any application, and release the key to insert the text.
I use the local Whisper Small model with the language set to English (India). For my speech, it is noticeably more accurate than Parakeet in VoiceInk. I make fewer corrections, and the result is more often what I intended to say.
The wait is also very noticeable. It can take a couple of seconds for the text to appear, sometimes longer depending on how much I recorded. Moving directly from VoiceInk on the M2 to Vocalinux on this laptop makes the difference impossible to miss.
Even so, it is still faster than I type.
A small improvement from the MX150
Vocalinux detected the MX150 in its settings, so I initially assumed it was already using the GPU. It was not. The installed whisper.cpp backend had no GPU support and was running on the CPU, even though the NVIDIA driver and Vulkan were working normally.
After rebuilding the backend with Vulkan enabled, Vocalinux could use the discrete GPU. Transcription became slightly faster, but the delay did not disappear.
The MX150 helps. It just does not make an older laptop feel like an M2 Mac.
The delay remains noticeable, but it has not stopped me using Vocalinux. Waiting a few seconds is an acceptable trade-off when the transcription needs fewer corrections and still reaches the screen faster than I could type it.