whisper-large-v3-urdu-ct2

CTranslate2-int8 conversion of kingabzpro/whisper-large-v3-turbo-urdu for fast CPU inference.

---

Installation

pip install faster-whisper ctranslate2

• faster-whisper provides a simple Python API on top of CTranslate2.
• ctranslate2 is the optimized inference backend.

---

Download

from faster_whisper import WhisperModel
model = WhisperModel("kingabzpro/whisper-large-v3-urdu-ct2", device="cpu", compute_type="int8")

---

Usage

from faster_whisper import WhisperModel

# Load the int8-quantized CTranslate2 model
model = WhisperModel(
    "kingabzpro/whisper-large-v3-urdu-ct2",
    device="cpu",         # run on CPU
    compute_type="int8"   # int8 quantization
)

# Transcribe a WAV, MP3 or NumPy array
segments, info = model.transcribe(
    "path/to/audio.wav",
    beam_size=5,          # ↔ speed vs. accuracy
    vad_filter=True       # skip silence for extra speed
)

# Join segments into a single string
text = " ".join(segment.text for segment in segments)
print(text)

---

Performance

• 5×–10× faster than the default PyTorch pipeline on a modern CPU
• Int8 weights reduce memory footprint by ~4×

---

Troubleshooting

• “ModuleNotFoundError” → check that faster-whisper and ctranslate2 are installed
• Noise or dropouts → try lowering beam_size or disabling vad_filter
• Out-of-memory → ensure no other heavy processes are running, or switch to float16 quantization