A level up on voice interaction with agents – and local(narrato.tech) |
A level up on voice interaction with agents – and local(narrato.tech) |
I'm excited about the tool so I wanted to share here, any feedback at all would be amazing :)
Key points: Dictate - press option + space to pop up the dictation tool, ramble while you work and the tool cleans up your text and pastes it into any selected text field (and the clipboard) Read Aloud - Control-Option-Space to read selected text, or the clipboard, aloud Clipboard Context - copy an error, issue, screenshot, or file while dictating; Narrato inserts it into the finished transcript where you were talking about it. I built this primarily for coding-agent prompts Local clipboard history - for copied text, images, and file references.
Dictation uses Parakeet TDT 0.6B v3 (4-bit, ~607 MB) and Read Aloud uses Kokoro 82M. Both are top of the range open source models in the field. Speech, transcripts, cleanup, and generated audio stay on the Mac. There is 0 cloud inference.
First model setup is about 1.5 GB; models are cached afterwards. Apple Silicon / macOS 14+ only; 16 GB memory recommended. Warmed M3 Pro / 16 GB development measurements: Read Aloud starts in roughly 0.6–0.7 s; a one-minute dictation reaches final text about 1.2 s after release. Those are one-machine development figures, not a broad benchmark. One time purchase no subscription; it's in beta with a 7-day free trial, then a $29.99 licence for 2 Macs.
I’d love feedback from people using Codex, Claude Code, Cursor, or similar tools: is Clipboard Context useful, and what would make you hesitant to use it or Narrato in general? I've only just put it out so I will be rolling out improvements and features frequently.
You can see it in action here if you're curious: https://www.youtube.com/@narratotech