|
|
--- |
|
|
language: |
|
|
- en |
|
|
- zh |
|
|
- de |
|
|
- es |
|
|
- ru |
|
|
- fr |
|
|
- ja |
|
|
- ko |
|
|
- pt |
|
|
- tr |
|
|
- pl |
|
|
- it |
|
|
- nl |
|
|
- sv |
|
|
tags: |
|
|
- whisper |
|
|
- openvino |
|
|
- int8 |
|
|
- intel-igpu |
|
|
- speech-recognition |
|
|
- automatic-speech-recognition |
|
|
- unicorn-amanuensis |
|
|
license: apache-2.0 |
|
|
pipeline_tag: automatic-speech-recognition |
|
|
--- |
|
|
|
|
|
# Whisper Base INT8 - Optimized for Intel iGPU 🚀 |
|
|
|
|
|
This is an **INT8 quantized** version of OpenAI's Whisper base model, specifically optimized for **Intel integrated GPUs**. |
|
|
|
|
|
## 🎯 Key Features |
|
|
|
|
|
- **4x smaller** than FP32 (75MB vs 280MB) |
|
|
- **2-4x faster inference** on Intel iGPU |
|
|
- **INT8 asymmetric quantization** |
|
|
- **100% weights quantized** to INT8 |
|
|
- **OpenVINO 2024.0+** compatible |
|
|
|
|
|
## 📊 Performance |
|
|
|
|
|
| Metric | Original | INT8 | Improvement | |
|
|
|--------|----------|------|-------------| |
|
|
| Model Size | 280MB | 75MB | **3.7x smaller** | |
|
|
| Inference Speed | 1.0x | 2-4x | **2-4x faster** | |
|
|
| Memory Bandwidth | 100% | 30-50% | **50-70% reduction** | |
|
|
|
|
|
## 🎮 Optimized for Intel Hardware |
|
|
|
|
|
- Intel Arc Graphics (A770, A750, A380) |
|
|
- Intel Iris Xe Graphics (12th Gen+) |
|
|
- Intel UHD Graphics (11th Gen+) |
|
|
|
|
|
## 📄 License |
|
|
|
|
|
Apache 2.0 |
|
|
|
|
|
## 🦄 Part of Unicorn Amanuensis |
|
|
|
|
|
Professional STT suite: https://github.com/Unicorn-Commander/Unicorn-Amanuensis |
|
|
|
