Voice2Text AI

Fast native Linux dictation with local AI. Record from your microphone, have speech transcribed on-device with Faster Whisper, prompt a local Ollama model, and hear the answer spoken back with a natural neural voice and an offline eSpeak NG fallback.

Built with GTK 4 and libadwaita, powered by GStreamer, and distributed as a Flatpak. Works on Wayland and X11 with PipeWire or PulseAudio.

Features

• On-device dictation — microphone audio is transcribed locally by Faster Whisper (nothing leaves your machine).
• Local AI prompts — send the transcript to any Ollama model you have pulled, with streaming responses.
• Natural speech output — Edge TTS voices stream immediately; if the network voice is unavailable, eSpeak NG speaks offline automatically.
• Adaptive segmentation — speech is split at pauses, so dictation flows naturally while transcribing in the background.
• Coordinated appearance — follows the system light/dark setting, with an explicit override in Preferences.

Install (Flatpak)

The easiest way is the Flatpak from the releases:

flatpak install --user Voice2Text-AI.flatpak
flatpak run io.github.crhy.voice2textai

For local AI, install Ollama and pull a model, for example:

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2:0.5b

qwen2:0.5b is the smallest useful model at ~350mb. Depending on hardware, you may want to upgrade to: ollama pull qwen3.6:27b *Note this model will run much better on a GPU with > 17gb of ram

The Whisper model downloads on first launch (the base model is the default; smaller models use less memory and start faster).

Run from source

Requires Python 3.11+, GTK 4, libadwaita, and GStreamer with the Python bindings. Install the Python dependencies and launch:

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python voice2text_ai.py

voice2text_ai.py is a thin launcher over the voice2text package; the PyPI entry point is voice2text-ai.

Keyboard shortcuts

Shortcut	Action
Ctrl+R	Start or stop dictation
Ctrl+Enter	Ask AI
Ctrl+Shift+C	Copy transcript
Ctrl+L	Clear
Ctrl+,	Preferences
Ctrl+Q	Quit

Building the Flatpak

The GitHub Actions workflow generates pinned Python dependencies, builds the Flatpak bundle, and attaches it to a draft release for every v* tag. To build locally, generate the pinned module and run flatpak-builder:

git clone https://github.com/flatpak/flatpak-builder-tools.git /tmp/flatpak-builder-tools
python3 /tmp/flatpak-builder-tools/pip/flatpak-pip-generator \
  --requirements-file=packaging/flatpak/requirements.txt \
  --runtime org.gnome.Sdk//48 \
  --prefer-wheels=ctranslate2,onnxruntime,tokenizers,av,numpy,pyyaml,protobuf \
  --wheel-arches=x86_64 \
  --output=python3-requirements-flatpak
flatpak-builder --user --install --force-clean build-dir io.github.crhy.voice2textai.yml

Configuration

Settings are stored in $XDG_CONFIG_HOME/voice2text-ai/config.json and edited from the Preferences dialog.

Acknowledgements

• Faster Whisper for on-device transcription
• Ollama for local language models
• Edge TTS for natural voices
• eSpeak NG for offline speech
• GTK 4, libadwaita, and GStreamer

License

MIT — see LICENSE.