Whisper-Small-MN-int8 (Faster-Whisper)

This model is an INT8 quantized Faster-Whisper version of openai/whisper-small,
converted from MinaNasser/Whisper-Small-MN using CTranslate2.

It is optimized for fast Arabic speech recognition with low memory usage.

---

Performance

• Loss: 0.365
• WER: 37.8043
• CER: 22.2604

---

Quantization Details

Setting	Value
Runtime	ctranslate2
Compute type	int8
Optimization	faster-whisper
Conversion tool	ctranslate2

---

Training Data

• Dataset: Arabic_STT_DS_AI_Transcriped
• Language: Arabic

---

Intended Use

• Fast Arabic ASR
• Real-time transcription
• CPU & low-VRAM GPU inference
• Edge & production deployment

---

Limitations

• INT8 quantization may slightly reduce accuracy
• Trained on a single Arabic dataset
• Dialect generalization is not guaranteed

---

Inference Example (Faster-Whisper)

from faster_whisper import WhisperModel

model = WhisperModel(
    "MinaNasser/Whisper-Small-MN-int8",
    compute_type="int8"
)

segments, info = model.transcribe("audio.wav", language="ar")

for segment in segments:
    print(segment.text)