PyScribe captures audio from Windows loopback devices (system audio/speakers) and uses OpenAI's Whisper model to perform speech recognition locally.

What it can do:

• Generates .wav files from system audio output
• Transcribes recordings using OpenAI's Whisper model running locally
• Works without internet connection or API keys

• Documentation: Record and transcribe meetings, lectures, or client calls to create searchable text archives without manual note-taking
• Content Creation: Generate subtitles, blog posts, or searchable libraries from podcasts, videos, or educational content
• Audio Archives: Convert existing audio recordings into searchable text for better organization and accessibility

git clone https://github.com/tobiaswuerth/pyscribe.git
cd pyscribe
py -m venv .venv
.\.venv\Scripts\activate
python.exe -m pip install --upgrade pip

# For CUDA-enabled GPUs
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128 

pip3 install -r .\requirements.txt

Additionally, you may also need to install ffmpeg and add the ffmpeg.exe to the OS PATH variable.

Adjust config.yaml before running.

py .\main.py

This starts two processes:

1. Recorder

   • Records system audio from the configured/selected loopback device
   • Saves recordings as .wav files in segments of configured length (default 5min)

2. Transcriber

   • Monitors output directory for untranscribed audio files
   • Processes files and saves transcription as .wav.txt

in the end, by default (depending on config), all individual segments of .wav and .txt files will be combined into one .wav and .txt file respectively, and the individual segment files are deleted unless otherwise configured.

• Recording and transcribing simultaneously works well for most scenarios
• For highest quality transcription (particularly for longer sessions), consider:
  1. Recording first (using py .\record.py)
  2. Transcribing afterwards (using py .\transcribe.py)

This sequential approach produces better results because the ASR model has more context and complete sentences when processing the entire audio file, resulting in better transcription quality overall.

For recording only:

py .\record.py

For transcription only:

py .\transcribe.py

For combining of individual files only:

py .\combine.py

• The recorder captures system audio output only (loopback devices), not microphone input
• Transcription quality depends on the Whisper model selected in the config file
• Recording is saved in WAV format with the device's native sample rate