raj7676/ivy-dictation

Ivy Dictation

The speech model inside Ivy, a free offline dictation app for Windows and Linux. Hold a key, talk, let go: clean text lands wherever your cursor is.

Ivy writes what you meant, not just what you said. People change their minds mid-sentence, and this model applies those corrections itself, in a single pass, with no second AI rewriting the text:

You say	Ivy types
"Let's meet on Tuesday, sorry, I mean Thursday."	Let's meet on Thursday.
"Book a table for four, actually make it six."	Book a table for six.
"Tell Sam the call is at 3, no wait, 4 PM."	Tell Sam the call is at 4 PM.
"Um, so, I I think we should, uh, ship it."	I think we should ship it.

It also drops fillers and stutters and adds punctuation and capital letters.

Get the app

You don't need to download anything from this page to use Ivy. The app downloads this model once on first start, then works fully offline.

• Windows: Microsoft Store or the setup file on GitHub

• Linux: open a terminal and paste the line for your system:

  • Ubuntu 22.04+, Linux Mint 21+, Debian 12: wget https://github.com/raj-7676/ivy-dictation/releases/download/v0.2.9/Ivy_0.2.9_amd64.deb && sudo apt install ./Ivy_0.2.9_amd64.deb
  • Fedora 37+: curl -LO https://github.com/raj-7676/ivy-dictation/releases/download/v0.2.9/Ivy-0.2.9-1.x86_64.rpm && sudo dnf install ./Ivy-0.2.9-1.x86_64.rpm

  Then open Ivy from your apps menu and hold Ctrl + Shift to talk. Ivy works fully on X11 desktops (Linux Mint, or "Ubuntu on Xorg" at the login screen). On Wayland, the default on Ubuntu and Fedora, Ctrl + Shift is the only hotkey and support is still being tested. Not available yet for Arch, ARM or 32-bit PCs. Full Linux guide.

• Source code: github.com/raj-7676/ivy-dictation

Files

One model, stored as two files (GGUF for llama.cpp):

File	What it is	Size
ivy-Q8_0.gguf	Language model, Q8_0	1.83 GB
mmproj-ivy-f16.gguf	Audio encoder and projector, F16	0.64 GB

How it was made

• Base: Qwen3-ASR-1.7B by the Qwen team, Alibaba Cloud.
• Fine-tune: LoRA on many hours of real people speaking (noisy rooms, cheap microphones, many accents) with human-written transcripts, plus thousands of scripted dictation paragraphs full of self-corrections, read in many voices over real background noise. Every training target was written by humans; no AI-generated transcripts.
• The LoRA was merged into the weights and converted to GGUF.

Speed

Measured in the Ivy app (llama.cpp, median of runs) on an RTX 4060 Laptop GPU and an Intel Core i7:

Clip length	GPU (Vulkan)	CPU only
5 s	0.22 s	1.95 s
15 s	0.33 s	3.20 s
30 s	0.34 s	4.88 s

Prompt format

Ivy feeds 16 kHz mono audio through llama.cpp's multimodal (mtmd) API with this prompt, where <__media__> is the audio:

<|im_start|>system
Write what the speaker means, ready to paste: apply their own corrections, keep every other word.<|im_end|>
<|im_start|>user
<__media__><|im_end|>
<|im_start|>assistant
language English<asr_text>

The model then writes the clean text and stops at <|im_end|>. It was trained on English.

License

Free to use, study and fine-tune; not for sale. The fine-tune is under the Apache License 2.0 with the Commons Clause condition (LICENSE): you may not sell a product or service whose value comes mainly from this model. The original Qwen3-ASR-1.7B stays under the plain Apache License 2.0 from its authors (Apache-2.0.txt).

Made by Glow.co.