whisper-small-urdu-int8-ct2

INT8 Quantized Version of khawajaaliarshad/whisper-small-urdu
Converted using CTranslate2 for faster inference on CPU/GPU, optimized for mobile and edge deployment.

---

Original Model

• Name: whisper-small-urdu
• Author: Khawaja Ali Arshad
• License: Apache 2.0
• This repository contains a quantized INT8 version; no retraining was performed.

---

Conversion Details

• Conversion tool: CTranslate2
• Quantization: INT8
• Purpose: Reduce model size and accelerate inference for low-resource environments (mobile/CPU/GPU).
• Backend support: Faster-Whisper compatible.

---

Benchmark (T4 GPU, 5-second audio sample)

Version	Inference Time	Notes
Original FP32	9.07 seconds	Standard Hugging Face PyTorch model
INT8 Quantized	0.54 seconds	Using CTranslate2, 16x speed-up

• Model size: ~967 MB original → INT8 version smaller (approx 1/4–1/3 the size; check actual folder size)

⚠️ Note: Actual speed may vary depending on device, CPU/GPU, and batch size.

---

Usage Example

from faster_whisper import WhisperModel

# Load the INT8 quantized model
model = WhisperModel(
    "whisper-small-urdu-int8-ct2",
    device="cpu",        # or "cuda" for GPU
    compute_type="int8"
)

# Transcribe an audio file
segments, info = model.transcribe("audio.wav")

print("Detected language:", info.language)
for segment in segments:
    print(segment.text)

License & Attribution

• This model is released under Apache 2.0 License, same as the original model.

• Original model by Khawaja Ali Arshad: Original Model

• INT8 quantization done by Muhammad Khubaib Ahmad using CTranslate2.

• Please retain proper attribution when using this model.

Recommendations

• For mobile/edge: Use INT8 version for faster inference and lower memory usage.

• For training/fine-tuning: Use the original FP32 model; quantized INT8 is not suitable for further training.

• For benchmarking: Test on your target hardware for accurate latency measurements.

• Compatibility: Fully compatible with faster-whisper API.