Nota 0.3.0
20 downloads
Nota 0.3.0: offline Amharic notes you can write or speak. This release builds Hohe, the Amharic speech model, into the app. On Apple silicon Macs with macOS 14 or later, Windows and Linux, recordings turn into text and dictation types as you speak, offline from the first launch, with no Python and no model to download (Windows may need one Microsoft component: see First launch).
What's new
- Hohe built in. Hohe now comes inside Nota and runs on your own computer, offline: record, dictate or transcribe a file straight away.
- Sentence ends. Where Hohe is built in, a transcribed recording gets full stops and question marks from a small punctuation model that ships with it. Dictation types the words as Hohe hears them, without marks.
- Accuracy. On a standard Amharic test (the first 150 clips of FLEURS), built-in Hohe gets about 7 characters in 100 wrong (6.82%), the same as Hohe's 8-bit model under Python. In an earlier run on those clips, that 8-bit model scored 6.89% and the full-size model 6.88%.
- Speed. On an M1 Pro, this release turned a 77-second recording into text, sentence ends included, in 35 seconds.
- Memory. Hohe needs about 3 GB of free memory while it transcribes a long recording, and gives it back five minutes after it was last used.
Upgrading from 0.2.0. If you set up Hohe through Python, Nota now uses the built-in copy instead, except on macOS 13, where it keeps using the Python setup. On Apple silicon the Python setup ran Hohe on the GPU, so transcription there may now be slower. The Python setup is still used for BuzzASR. Your notes stay as they are.
Downloads
| System | Files |
|---|---|
| macOS 13 or later, Apple silicon | Nota-0.3.0-mac-arm64.dmg or Nota-0.3.0-mac-arm64.zip (on macOS 13, speech needs the setup under macOS 13 below) |
| macOS 13 or later, Intel | Nota-0.3.0-mac-x64.dmg or Nota-0.3.0-mac-x64.zip (no transcription or dictation: no speech model runs on Intel Macs) |
| Windows, x64 (Intel or AMD) | Nota-0.3.0-win-x64-setup.exe, or the portable Nota-0.3.0-win-x64.zip |
| Windows, ARM64 | Nota-0.3.0-win-arm64-setup.exe, or the portable Nota-0.3.0-win-arm64.zip |
| Linux, x86-64 (Intel or AMD) | Nota-0.3.0-linux-x86_64.AppImage or Nota-0.3.0-linux-x64.tar.gz |
| Linux, ARM64 | Nota-0.3.0-linux-arm64.AppImage or Nota-0.3.0-linux-arm64.tar.gz |
Not sure which? On a Mac, Apple menu → About This Mac: "Chip: Apple M…" is Apple silicon, "Processor: … Intel" is Intel. On Windows 11, Settings → System → About → System type says x64-based or ARM-based. On Linux, uname -m prints x86_64 or aarch64.
Take the first file in your row: the .dmg (open it and drag Nota to Applications), the -setup.exe (installs Nota) or the .AppImage (one file you run). The .zip and .tar.gz hold the same app unpacked: on a Mac drag Nota from it to Applications, on Windows open Nota.exe inside it, on Linux run nota inside it.
The Apple silicon, Windows and Linux files are about 0.9 GB each (GitHub lists them as 884–933 MB) because the speech models ship inside them. The Intel Mac files leave the models out and are about 125 MB.
First launch
- macOS: Nota isn't notarised by Apple yet, so macOS blocks its first launch. Drag Nota to Applications, then either open Terminal (Applications → Utilities), paste
xattr -dr com.apple.quarantine /Applications/Nota.appand press Return; or open Nota once, and when macOS blocks it, go to System Settings → Privacy & Security, choose Open Anyway, then open Nota again and confirm. - Windows: Nota isn't signed yet. Your browser may warn that the file isn't commonly downloaded: choose to keep it. When Windows says "Windows protected your PC", choose More info → Run anyway. Built-in Hohe also needs the Microsoft Visual C++ Redistributable: on that page take
vc_redist.x64.exe(on ARM64 PCs,vc_redist.arm64.exe). The setup guide says Hohe is ready even without it; if Nota then says "Hohe couldn't start" ("Hohe መጀመር አልቻለም" in Amharic) when you record, install it, then close Nota and open it again. - Linux: AppImages need FUSE 2, which Ubuntu 22.04 and later don't install. On Ubuntu 24.04 run
sudo apt install libfuse2t64once (on 22.04,libfuse2). Then, in a terminal in your Downloads folder, runchmod +x Nota-0.3.0-linux-x86_64.AppImageand./Nota-0.3.0-linux-x86_64.AppImage(on ARM64, use thelinux-arm64file). If it closes at once, as Ubuntu 24.04's sandbox rules can make it do, add--no-sandboxto the second command. - Then: Nota opens its setup guide. Its Offline speech page should say "Hohe is ready. Speech works offline.": there is nothing to set up. Allow the microphone when asked, the first time you record or dictate. For the interface in Amharic, open Settings and choose አማርኛ under Language · ቋንቋ.
macOS 13
Built-in Hohe needs macOS 14 or later. On Apple silicon Macs still on macOS 13, Hohe runs through Python as in 0.2.0, and transcribed recordings don't get sentence ends.
- Set up Hohe in 0.2.0? Nothing to do: install 0.3.0 over the old Nota, and it keeps using your setup.
- New to Nota? The setup guide's Set up Hohe button downloads the model (2.4 GB) and a private Python runtime (about 0.6 GB) once. It needs Python 3.10–3.13 or uv installed first, and about 4 GB of free disk. Guide: Setting up Hohe on macOS 13.
Known issues
- On Intel Macs the setup guide still offers Set up Hohe, but the setup can't finish there: skip that page. Recording still saves the sound, but the note then says Hohe isn't set up, which can't be fixed on an Intel Mac.
Tested
These files were built on an Apple silicon Mac. The Apple silicon DMG above was run there: the setup guide showed Hohe built in and ready, and it transcribed a recording with sentence ends. The Intel Mac, Windows and Linux builds have not been run on those systems; the unpacked Windows and Linux builds these files were made from were checked to carry the right ONNX Runtime and the models. Please open an issue if anything goes wrong.
Credits
- Hohe ASR (
snapwre/hohe-asr-amharic) is by Chapi and the Dataset.ET contributors, licensed CC BY 4.0, fine-tuned frombadrex/Ethio-ASR-multilingual-600M(CC BY 4.0). Nota ships it with its weights quantized to 8 bits and exported to ONNX. - The punctuation model,
1-800-BAD-CODE/xlm-roberta_punctuation_fullstop_truecase, is licensed Apache 2.0 and built on XLM-RoBERTa (MIT). Nota ships it cut to Amharic and quantized to 8 bits. - ONNX Runtime is by Microsoft, MIT licensed.
Nota is MIT licensed.