Features

• Real-time speech-to-text via OpenAI Whisper (faster-whisper, CPU · int8 fully offline after first download).
• Brazilian Portuguese and English transcription.
• WebRTC VAD (primary) + energy-based fallback.
• Context-aware transcription (Previous sentences used as prompt for better accuracy).
• Compatible with Windows 10/11 and Linux (GTK3).
• All background threads are daemon threads — no process leaks on abnormal exit.
• Clipboard access uses wx.TheClipboard (OS-level, no third-party service).

Requirements

    pip install -r requirements.txt

• The Whisper base model (~150 MB) is downloaded from Hugging Face and cached locally, all subsequent runs are fully offline.

Required for wxPython on Linux (GTK3)

    sudo apt install libgtk-3-dev

Ubuntu/Debian pre-built wheel:

    pip install -f https://extras.wxpython.org/wxPython4/extras/linux/gtk3/ubuntu-22.04 wxPython

Running

    python main.py

AI / Audio Pipeline

        Microphone  (sounddevice, 16 kHz, 30 ms frames)
    │
    ▼
WebRTC VAD  ──► energy-based fallback if unavailable
    │  complete speech utterances only
    ▼
TranscriptionWorker queue  (background thread, thread-safe)
    │
    ▼
faster-whisper  (CPU · int8)
    ├─ Silero VAD filter        (removes residual silence)
    ├─ Context prompt           (rolling 5-sentence window → accuracy boost)
    ├─ Beam search (beam=5)     + temperature fallback
    └─ Confidence gate (≥0.35)  (discards low-quality results)
    │
    ▼
wx.CallAfter  ──► main thread ──► TextCtrl.AppendText()

Disclaimer

• This tool is under development and may contain bugs.