Shell-Game-Transcripts

Shell Game Transcripts — Troubleshooting

Summarisation fails for every episode

Almost always Ollama. Check both halves:

ollama list          # is the model pulled?
ollama pull llama3.1:8b
ollama ps            # is the server running?

Summarisation talks to Ollama over HTTP. If the server is down or the model is absent, every episode fails the same way rather than some succeeding.

ffmpeg not found during transcription

Whisper shells out to ffmpeg to decode audio. Install it and make sure it is on PATH — a Python package will not substitute.

CUDA out of memory

The Whisper model is set to turbo. Larger models need more VRAM. Either switch to a smaller one or let it fall back to CPU — Configuration has the line to edit.

Transcription is extremely slow

Expected on CPU. Install the CUDA build of PyTorch — see Installation. Note that PyTorch installed for CPU will silently stay on CPU even with a working GPU present; the index URL matters.

A stage says there is nothing to do

Its resume log already lists the items. That is the mechanism that makes long runs interruptible, and it also means edited settings do not trigger a rebuild. Delete the log to force one — Pipeline lists which log belongs to which stage.

Downloads fail or return nothing

Publishers move feeds and change URL formats. Update yt-dlp first, since it changes most often:

pip install --upgrade yt-dlp

If that does not help, the feed or repository the download stage targets has probably moved — see External Services.

LanguageTool fails to start

language-tool-python starts a local Java process. Without a Java runtime the cleanup stage cannot run; the earlier stages are unaffected, so transcripts and summaries still get produced.