screencastgen

Generate narrated audio

The audio pipeline extracts and cleans document text, splits it into chunks, synthesizes each chunk, and concatenates the results.

Local Qwen synthesis

screencastgen audio book.pdf --backend qwen --device cuda -o book.wav

Use a reference recording for voice cloning:

screencastgen audio book.pdf \
  --backend qwen \
  --model 1.7B \
  --ref-audio voice.wav \
  --ref-text "Transcript of the reference recording."

--ref-text can be omitted when the selected runtime can obtain the transcript through the surrounding workflow. Supplying an accurate transcript directly is the most predictable CLI path.

Remote synthesis

screencastgen audio book.pdf \
  --backend remote \
  --tts-server-url http://gpu-vm:8100 \
  --tts-concurrency 4

Increase concurrency only when the remote server has enough workers and GPU capacity. See the remote GPU guide for server setup.

Resume or restart

Chunk outputs and state are retained under the output directory. Re-running the same command resumes incomplete work. Add --clean to synthesize everything again, or --no-concat to retain chunks without creating the final audio file.

For every option, run screencastgen audio --help or consult the CLI reference.

This site is open source. Improve this page.