Flash is a coding agent that runs on your machine, on your own models, through Ollama. It has had voice mode in the terminal for a while. As of 0.5.11, you can talk to its web UI too.
The whole loop is local
Most voice features ship your audio to a cloud speech API. My browser blocks that API anyway.
So the web UI does what the terminal's voice mode already did. The browser records, and Flash turns speech into text with the same Vosk model and reads replies back with the same Piper voice. Both run on your computer. Nothing leaves it.
No hands
Click the voice button once. After that, the words the terminal already knew work here too, plus a couple of new ones:
- Talk, then pause. Flash sends it.
- Say "interrupt" to cut it off mid-sentence.
- Say "pause", then "resume".
- Say "stop voice" to finish.
The hard part was permission prompts. Flash asks before it runs a command, and a click would break the whole idea. So it reads the question out, command included, and you answer "yes" or "no".
It reads everything
The whole reply, sentence by sentence, with the current one highlighted and scrolled into view. Code blocks get skipped, and so do emojis. Voice turns also tell the model to keep it short.
The bug that was funny
Early on it would sometimes speak gibberish. It turned out that anything inside [[ ]] gets read by the speech engine as raw sound codes. Code like is_expired >= now went in character by character. Now symbols become words: "is expired at least now".
Try it
curl -fsSL https://raw.githubusercontent.com/Natuworkguy/Flash/main/install.sh | bash
Release notes: https://github.com/Natuworkguy/Flash/releases/tag/v0.5.11
Top comments (0)