Hi everyone,
Ahead of version 1.4, which will rework the “meeting” feature, here is a new maintenance release of Dictée, focused on responsiveness and reliability. The highlights:
New Parakeet Int8 model, more responsive on CPU than Parakeet fp32
The compact model now runs where it is fastest, for better performance out of the box. Measured on an i7-13700H / 64 GB RAM, a clean ~15 min audio reading (LibriVox):
Comparison on CPU (word error rate, lower = better):
| Model (CPU) | Word error rate (WER) | Processing time (~15 min audio) | Memory |
|---|---|---|---|
| FP32 (reference) | 8.0% | ~76 s | ~2.5 GB |
| Int8 (new default) | 8.4% | ~52 s | ~1.2 GB |
Parakeet int8 is ~32% faster and ~1.4 GB lighter, for near-identical accuracy (+0.4 pt error). On very short clips the gap can reach ~+2.5 pt, but in continuous dictation (the real-world use) it stays minimal.
Switching back to FP32 is available in the settings.
Here is how to set up Parakeet int8 from Dictée’s configuration window
A few fixes:
-
Push-to-talk input fixed (issue #8) — the last character no longer repeats after a while of dictation, on multi-keyboard setups or under Wayland.
-
Works with keyboard remapping tools (issue #10) — logiops, keyd, kanata… can now trigger dictation, via a new option in the settings.
-
More robust day-to-day: Verified model downloads, more reliable Whisper (better CPU/GPU choice, fewer made-up words); lighter desktop widget at idle; settings better preserved; broader Fedora compatibility and more stable speaker diarization.
Note for KDE Neon users
*KDE Neon ships a Qt version that is ahead of the Python component Dictée uses for audio, PyQt6 — specifically its QtMultimedia module (the python3-pyqt6.qtmultimedia package is at 6.10 while qt6-multimedia is at 6.11).
1.3.5 fixes the crash this caused (issue #13): the app starts and works normally, and both dictation and transcription work (audio goes through PipeWire). The remaining Neon-specific limitation is therefore in:
the built-in audio player of the transcription window.
the mic-test level meter in the configuration windows and the wizard.
This is a KDE Neon packaging quirk (independent of Dictée) — other Plasma 6 distributions are not affected. Everything should return to normal once the package alignment is fixed on the KDE Neon side.
Download: Release v1.3.5 · rcspam/dictee · GitHub
Available packages: Debian / Ubuntu (.deb), Fedora / openSUSE (.rpm), Arch Linux (PKGBUILD), universal archive (.tar.gz), and the standalone .plasmoid widget.
Installation:
curl -fsSL https://raw.githubusercontent.com/rcspam/dictee/master/install.sh | bash
As a reminder, Dictée is a 100% local voice dictation app for Linux, integrated with Plasma 6: speech recognition in 25 languages, meeting transcription with speaker labeling, and optional grammar correction by a local LLM — all offline by default.
Your feedback is welcome: it feeds directly into the upcoming v1.4.