Talk2Text

Talk2Text is a local Python app for microphone transcription. It records audio with Qt, transcribes it locally with faster-whisper, and can optionally send the finished transcript to a local Ollama model for cleanup.

This project is an early desktop MVP.

What works today:

• microphone recording through Qt Multimedia
• hold-to-record with Space while the main window is focused
• local Whisper transcription
• optional live transcription
• manual Ollama transcript polish
• transcript history persisted locally

What is still rough:

• no persistent settings storage yet
• no global push-to-talk yet
• limited automated coverage around GUI behavior

microphone -> Qt audio capture -> faster-whisper -> transcript
                                             \
                                              -> optional Ollama polish

• uv
• local Ollama installation if you want transcript polish

The project targets Python 3.12. uv can use an existing Python or manage Python for you.

For GPU acceleration, a working NVIDIA CUDA 12 + cuDNN 9 stack is recommended. The app can fall back to CPU transcription.

uv sync
uv run talk2text

Alternative launch command:

uv run python -m talk2text

Install uv with the official installer:

curl -LsSf https://astral.sh/uv/install.sh | sh

Other installation methods are available in the uv docs.

• Whisper model: turbo
• Ollama model: qwen3:8b
• Live transcription: disabled by default
• Ollama polish: disabled by default

export TALK2TEXT_WHISPER_MODEL=large-v3
export TALK2TEXT_OLLAMA_MODEL=qwen3:8b
export TALK2TEXT_OLLAMA_BASE_URL=http://localhost:11434
export TALK2TEXT_LANGUAGE=
export TALK2TEXT_SAMPLE_RATE=16000
export TALK2TEXT_ENHANCE_WITH_OLLAMA=0
export TALK2TEXT_LIVE_TRANSCRIPTION=0

Leave TALK2TEXT_LANGUAGE empty for language auto-detection.

Sync the project environment:

uv sync

Run checks:

uv run python -m compileall src tests
uv run python -m unittest discover -s tests
uv run ruff check src tests

When dependencies change, refresh the lockfile with:

uv lock

src/talk2text/
  audio.py           Qt microphone capture
  transcription.py   Whisper integration
  ollama_client.py   Ollama cleanup client
  pipeline.py        transcription pipeline
  ui.py              desktop UI

• First Whisper use may download the selected model.
• Ollama polish requires a local Ollama server running at http://localhost:11434 unless you override it.
• If live transcription feels heavy, keep it disabled and use record-then-transcribe mode.

See CONTRIBUTING.md.

Apache-2.0. See LICENSE and NOTICE.