Ogaal ASR CTranslate2 INT8
CPU-optimized INT8 conversion of
Ogaal-Labs/Ogaal-ASR for
use with faster-whisper.
The original model is a Somali fine-tuned Whisper large-v3 checkpoint. This conversion reduces memory use and improves CPU inference speed while keeping the same tokenizer and Somali task configuration.
from faster_whisper import WhisperModel
model = WhisperModel(
"Ogaal-Labs/Ogaal-ASR-CT2-INT8",
device="cpu",
compute_type="int8",
cpu_threads=4,
)
segments, _ = model.transcribe(
"audio.wav",
language="so",
task="transcribe",
beam_size=1,
condition_on_previous_text=False,
)
text = " ".join(segment.text.strip() for segment in segments)
See the original model card for training data, evaluation metrics, scope, and limitations.