API Reference
talk_tag.api
talk_tag.api
StartupContext
dataclass
Startup metadata emitted by model pull and annotation startup callbacks.
Backend semantics:
- cuda / mps / cpu: backend was verified by loading a
TransformersAnnotator.
- external: backend was not verified by talk-tag. This is used for both:
- caller-managed engines (annotate_path(..., engine=...)), and
- unverified pulls (pull_model(..., verify_load=False)).
Use model_source to disambiguate these two cases:
- external_engine => caller supplied the engine.
- any repository/local model source => backend is unknown/unverified.
annotate_path(input_path, output_dir, target_speaker, *, investigator_speaker=None, device='auto', hf_cache_dir=None, granularity='standard', error_tags=None, limit=0, show_target=False, print_debug_lines=False, speaker_field=None, text_field=None, case_insensitive_speaker=False, continue_on_error=True, show_progress=True, engine=None, startup_callback=None)
Run annotation over a single file or a folder.
If engine is provided, startup metadata reports backend="external" and
model_source="external_engine" because talk-tag did not resolve/load the
model runtime itself.
pull_model(*, hf_cache_dir=None, device='auto', verify_load=True)
Download/resolve model assets and optionally verify model loading.
When verify_load=False, the returned context uses backend="external"
to indicate an unverified/unknown runtime backend (not a caller-supplied
engine). In that case, model_source still reflects the resolved source.