Shell-Game-Transcripts

Shell Game Transcripts — Configuration

There are no environment variables, config files, or command-line flags. Every setting below is a literal inside a stage script, which is precisely why it is worth listing — changing behaviour means editing code, and knowing which line saves you reading all of it.

Settings

Setting Current value Defined in Why it matters
Whisper model turbo 3_transcriber.py Accuracy against transcription speed and VRAM
Transcription language en 3_transcriber.py Whisper is pinned to this language rather than detecting it
Ollama model llama3.1:8b 4_summarizer.py Summary quality; must already be pulled in Ollama
Summary chunk size 2000 tokens 4_summarizer.py How much context each summarisation pass sees
Chunking tokenizer bert-base-uncased 4_summarizer.py Only used to measure chunk length, not to generate text
Grammar locale en-CA 5_cleanup.py Which spelling conventions the cleanup stage enforces

Notes on the ones that bite

Whisper model (turbo). Larger models transcribe more accurately and more slowly, and need more VRAM. If transcription fails with an out-of-memory error, this is the line to change in 3_transcriber.py.

Ollama model (llama3.1:8b). Must be pulled in Ollama before the stage runs — ollama pull it, or summarisation fails for every episode rather than degrading.

Language (en). Whisper is pinned rather than auto-detecting. That is faster and more reliable for a single-language show, and wrong for anything multilingual.

Grammar locale (en-CA). Decides which spellings the cleanup stage treats as errors. Changing it rewrites the corrected transcripts accordingly.

Chunk size (2000 tokens). Transcripts are split into chunks, each summarised separately, and the summaries then combined. Larger chunks give the model more context per pass but risk exceeding what it handles well.

Changing a setting safely

Stages skip work already recorded in their log, so editing a model and re-running does not regenerate existing output. Delete the relevant log first — see Pipeline.