Skip to content

fix(model): support large whisper models and default to large-v3-turbo - #6

Merged
zeke merged 1 commit into
mainfrom
fix/large-whisper-models
Jun 17, 2026
Merged

zeke merged 1 commit into
mainfrom
fix/large-whisper-models

Conversation

@zeke

@zeke zeke commented Jun 17, 2026

Copy link
Copy Markdown
Owner

This PR fixes large whisper models being unusable and makes large-v3-turbo the default.

_model_short_name() feeds the model name to whisper-cli --dtw. whisper.cpp's DTW presets for large models are dotted (large.v3, large.v3.turbo), but the model filenames are hyphenated (ggml-large-v3-turbo.bin). The old code stripped the ggml- prefix and passed the hyphenated name through, so --dtw large-v3 was rejected and any large model failed. The fix converts large-* names to dotted form.

It also switches the default model from medium.en to large-v3-turbo and updates MODEL_SIZES, the download help text, tests, and docs to match.

Changes:

  • caption.py: dotted-preset conversion for large models; default model, MODEL_SIZES, help text, docstring
  • tests/test_caption.py: assertions for large-v3 and large-v3-turbo mapping
  • AGENTS.md, README.md: model references and the dotted-preset gotcha

ruff and pytest pass locally.

whisper.cpp's --dtw presets for large models are dotted (large.v3,
large.v3.turbo) but the model filenames are hyphenated, so _model_short_name
produced invalid presets like 'large-v3' and whisper-cli rejected them.
Convert large-* names to dotted form.

Also switch the default model from medium.en to large-v3-turbo and update
MODEL_SIZES, download help, tests, and docs to match.
@zeke
zeke merged commit 3a6ddfd into main Jun 17, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant