nunu-willump/PhoWhisper-medium-ct2

PhoWhisper-medium — CTranslate2

FP16 CTranslate2 conversion of vinai/PhoWhisper-medium at revision 55a7e3eb6c906de891f8f06a107754427dd3be79. Converted and redistributed by nunu-willump for Chep. The original model authors retain credit; this repository does not add training or fine-tuning.

All five inference files are byte-for-byte identical to the conversion evaluated in Chep's Kaggle T4 pilot (six FLEURS test samples per language). See conversion.json for hashes and converter versions. This is a packaging equivalence check, not a new general accuracy claim.

Usage

from faster_whisper import WhisperModel
model = WhisperModel("nunu-willump/PhoWhisper-medium-ct2", device="cpu", compute_type="float32")
segments, info = model.transcribe("audio.wav", language="vi")
for segment in segments:
    print(segment.start, segment.end, segment.text)

GPU users can choose device="cuda", compute_type="float16". The saved weights are FP16; CTranslate2 selects the requested supported execution precision at load time. No PyTorch or Transformers installation is required for inference with faster-whisper.

Source and license

Read UPSTREAM_MODEL_CARD.md for the original authors, training data, citations and limitations. The source model declares bsd-3-clause. License and original Whisper copyright notices are included. For Spanish, the upstream card declares MIT without a separate fine-tuner copyright notice; the inherited Whisper MIT notice is retained.

PhoWhisper citation: Thanh-Thien Le, Linh The Nguyen and Dat Quoc Nguyen, "PhoWhisper: Automatic Speech Recognition for Vietnamese", ICLR 2024 Tiny Papers.