Whisper Turbo AZ for faster-whisper
This is the CTranslate2 fp16 export of hasan0v/whisper-turbo-az, an Azerbaijani fine-tune of Whisper large-v3-turbo. It is ready to use with faster-whisper, WhisperX and other CTranslate2 runtimes.
The scores in the table below were measured on this exact export: faster-whisper, fp16, beam 5.
| FLEURS-az test (921 clips) | WER ↓ | CER ↓ |
|---|---|---|
| LocalDoc/azerbaijani-whisper-turbo | 13.11% | 3.39% |
| Whisper Turbo AZ v2 (this model) | 10.90% | 2.86% |
On a laptop RTX 4070 it runs at a real-time factor of 0.035, which is about 28× faster than real time.
Usage
pip install faster-whisper
from faster_whisper import WhisperModel
model = WhisperModel("hasan0v/whisper-turbo-az-ct2", device="cuda", compute_type="float16")
# CPU: WhisperModel("hasan0v/whisper-turbo-az-ct2", device="cpu", compute_type="int8")
segments, info = model.transcribe(
"audio.wav",
language="az",
beam_size=5,
vad_filter=True,
condition_on_previous_text=True, # trained with previous-text context
)
for s in segments:
print(f"[{s.start:6.1f} → {s.end:6.1f}] {s.text}")
For fast batch transcription of long files:
from faster_whisper import BatchedInferencePipeline
pipe = BatchedInferencePipeline(model)
segments, _ = pipe.transcribe("lecture.mp3", language="az", batch_size=16)
print(" ".join(s.text.strip() for s in segments))
The training data, method, hyperparameters, limitations and licences are documented on the full model card.