Views
No views yet
bfloat16 (the base model is bf16-trained — converting to
float16 silently truncates multi-segment transcription, see faster-whisper#567). On
Ampere+ GPUs load with compute_type="bfloat16"; on GPUs without bf16 support (e.g.
T4) load with compute_type="float32" — widening bf16 → float32 is lossless.1ct2-transformers-converter \
2 --model primeline/whisper-large-v3-turbo-german \
3 --output_dir whisper-large-v3-turbo-german-ct2 \
4 --quantization bfloat16 \
5 --copy_files preprocessor_config.json
6# tokenizer.json generated separately so the model works fully offline (HF_HUB_OFFLINE=1)1from faster_whisper import WhisperModel
2
3model = WhisperModel("tnfru/whisper-large-v3-turbo-german-ct2",
4 device="cuda", compute_type="bfloat16")
5segments, info = model.transcribe("audio.ogg", language="de",
6 beam_size=15, best_of=5)
7print(" ".join(s.text for s in segments))