BrassBones/kb-whisper-large-subtitle-ct2

kb-whisper-large-subtitle (CTranslate2 / faster-whisper)

A CTranslate2 conversion of KBLab/kb-whisper-large at its subtitle revision — KBLab's condensed-output checkpoint — so it can be loaded directly by faster-whisper.

This is not an official KBLab release. All credit for the model belongs to KBLab (National Library of Sweden); this repository only contains a format conversion, published with their OK because the subtitle tag of the upstream repository ships transformers weights only. If KBLab publish an official CT2 version of this checkpoint, prefer theirs.

Provenance

Source repository	KBLab/kb-whisper-large
Source revision	subtitle tag, commit 50b62f493fa513926007d388f76cce9659bce123
Converter	ct2-transformers-converter (ctranslate2 4.8.1)
Stored precision	float16 (--quantization float16)
Files	config.json, model.bin, preprocessor_config.json, tokenizer.json, vocabulary.json
Verified with	faster-whisper 1.2.1, CPU, compute_type="int8"

Exact command:

ct2-transformers-converter --model KBLab/kb-whisper-large --revision 50b62f493fa513926007d388f76cce9659bce123 --output_dir kb-whisper-large-subtitle-ct2 --copy_files tokenizer.json preprocessor_config.json --quantization float16

Usage

from faster_whisper import WhisperModel

model = WhisperModel("BrassBones/kb-whisper-large-subtitle-ct2",
                     device="cpu", compute_type="int8")   # or device="cuda", compute_type="float16"
segments, info = model.transcribe("audio.wav", language="sv", beam_size=5)
for s in segments:
    print(f"[{s.start:.2f} -> {s.end:.2f}] {s.text}")

What the subtitle variant is

KBLab train KB-Whisper in stages and publish several output styles of the
large model. main (default) is the Stage 2 checkpoint; strict is
verbatim; subtitle is fine-tuned towards the shorter, condensed phrasing
used in broadcast subtitles. Because it intentionally drops filler and
compresses wording, it will score worse than main on verbatim WER
benchmarks — that is by design, not a conversion defect. See the upstream
model card for training details and evaluation.

License

Apache-2.0, inherited from the upstream model. Please cite KBLab's work when
you use it:

@misc{kb-whisper,
  title  = {KB-Whisper},
  author = {KBLab, National Library of Sweden},
  url    = {https://huggingface.co/KBLab/kb-whisper-large}
}