faster-whisper CT2: Whisper-small LoRA for Mandarin-English Lectures
CTranslate2 (float16) conversion of Creaturelove7/whisper-small-lora-ntu-ml2021 (LoRA merged into openai/whisper-small), ready for faster-whisper and speaches. See the adapter repository for results, training details, limitations, and data provenance.
from faster_whisper import WhisperModel
model = WhisperModel(
"Creaturelove7/faster-whisper-small-ntu-ml2021-ct2",
device="cuda", compute_type="int8",
)
segments, info = model.transcribe("audio.wav", language="zh", beam_size=5)
print("".join(segment.text for segment in segments))
On 11,919 held-out test utterances the underlying model reaches 8.51% CER versus 25.53% for base whisper-small. Output is normalized Simplified Chinese with English technical terms preserved; the behavior survives int8 quantization.