Local AI Weekly #10: Say It, Don't Type It Ubuntu 26.10 shipped Myna, Canonical's local speech-to-text tool that runs Parakeet, Whisper or FunASR on-device and needs no internet once the model is downloaded, according to Local AI Weekly #10. The newsletter also noted WhisPaste 1.5, released 16 November under the MIT license, added local transcription of audio and video files (MP3, FLAC, MKV, WebM) and encrypted history and notes, while Koe pairs whisper.cpp with a local Ollama model for transcript cleanup. The piece argues voice is the strongest case for local AI because speech models are small enough for an ordinary laptop and voice data is highly personal. I have said it in the past and I am saying it again. Voice is going to be the preferred input method in the future, be it smartphones or desktop computers. And Voice is one of the best cases for local AI. Speech models are small enough to run on an ordinary laptop. And your voice is about as personal as data gets. After all, you surely don't want every half-formed sentence you mutter going to someone else's server to train their AI. ๐Ÿง Linux gets a voice input Ubuntu 26.10 https://itsfoss.com/news/myna-ai-speech-to-text-tool/ shipped Myna, Canonical's speech-to-text tool. You press a shortcut, speak, and the text appears in whatever app you're using. It runs Parakeet, Whisper or FunASR locally and needs no internet once the model is downloaded. Myna is not the only such tool. There are many other tools like Vocalinux https://itsfoss.com/news/vocalinux-beta-release/ and TuxTalk https://github.com/ronb1964/TuxTalk?ref=itsfoss.com . Commercial tools like Superwhisper are also available for Linux but they are not really local. Most dictation tools, from Myna to Superwhisper, do the same four things. First, something decides when you're actually speaking and trims the silence. Then a speech model, usually Whisper or Parakeet, turns your audio into raw text. Some tools pass that text through a local LLM to fix punctuation, terms and formatting. Finally, the tool types the result into your active window. The cleanup step is optional, and it isn't free. A second model means more RAM in use and a longer wait before your text appears. For short messages, raw Whisper output is often good enough. ๐Ÿ” Discover AI tools WhisPaste 1.5 https://github.com/whispaste/whispaste?ref=itsfoss.com came out on 16 November. This MIT-licensed app transcribes on your machine with whisper.cpp or Parakeet, and cloud providers are optional. The new version can transcribe audio and video files https://whispaste.de/en/blog/neu-in-whispaste-1-5/?ref=itsfoss.com locally too, from MP3 and FLAC to MKV and WebM. Your history and notes are now stored encrypted. Koe https://github.com/shuhei0866/koe?ref=itsfoss.com is the one that does the LLM cleanup step locally. It's written in Rust and MIT-licensed. Speech goes through whisper.cpp, and the transcript is then tidied up by a model running in Ollama. The developer built it for writing Japanese on Ubuntu, but speech recognition covers any language Whisper does. ๐Ÿงช On my desk I use Proton's ecosystem at It's FOSS. I use their email servic https://go.getproton.me/aff c?offer id=37&aff id=1173&ref=itsfoss.com e for @itsfoss emails, their VPN, cloud storage and password manager. The entire Proton ecosystem is built on the privacy promise. Proton also offers its own private AI called Lumo https://go.getproton.me/aff c?offer id=68&aff id=1173&ref=itsfoss.com . Up until now, it was boring and wasn't of much use, it was just a chatbot, like what chatGPT was back in 2024. But yesterday, Proton revamped Lumo with a desktop app https://itsfoss.com/news/proton-lumo-desktop/ and it now has the options for skills, connectors and accessing system files and shell. And now it is actually useful. Your data stays private and yet you can use it for email triaging, GitHub management and several other tasks. It still has a long way to go, but at least Proton has taken the step in the right direction. ๐Ÿ—‚ AI jargon: VAD VAD stands for voice activity detection. It's a tiny model that listens to your microphone and decides whether a sound is speech or just noise, your fan or the keyboard. Dictation tools use it to cut silence before the audio reaches the speech model. Less audio means faster transcription. It also means Whisper doesn't get the chance to "hear" words in silence, which it's known to do. Silero VAD https://github.com/snakers4/silero-vad?ref=itsfoss.com is the common choice, and Koe supports it. โšก Quick Tip Most of these tools let you choose which Whisper model to download. Bigger isn't always better for dictation. If you only dictate in English, pick a model ending in .en , like base.en or small.en . These English-only versions tend to be more accurate than the multilingual model of the same size, especially at the smaller sizes. On a laptop without a good GPU, start with small.en and move up only if it keeps getting your words wrong. A model that's slow to respond will annoy you more than the occasional typo. See you next week ๐Ÿ™‚