Update README.md
Browse files
README.md
CHANGED
|
@@ -19,28 +19,20 @@ Optimized Whisper ONNX models packaged for easy deployment. Each zip contains al
|
|
| 19 |
|
| 20 |
| Model | Language | Size | Target Use | Download |
|
| 21 |
|-------|----------|------|------------|----------|
|
| 22 |
-
| **
|
| 23 |
-
| **Small
|
| 24 |
-
| **
|
| 25 |
-
| **Medium Multilingual** | 99 languages |
|
| 26 |
-
| **Large v3 Turbo** | 99 languages |
|
| 27 |
|
| 28 |
-
## Size Comparison vs GGML Q5_0
|
| 29 |
-
|
| 30 |
-
All models are **smaller** than equivalent GGML Q5_0 models:
|
| 31 |
-
|
| 32 |
-
- Medium English: 486MB vs 515MB GGML ✅ (-29MB)
|
| 33 |
-
- Small models: ~85-110MB vs 182MB GGML ✅ (-70-97MB)
|
| 34 |
-
- Large v3 Turbo: 530MB vs 574MB GGML ✅ (-44MB)
|
| 35 |
|
| 36 |
## Contents of Each Zip
|
| 37 |
|
| 38 |
-
Each zip file contains
|
| 39 |
|
| 40 |
### ONNX Model Files
|
| 41 |
- `encoder_model_quantized.onnx` - Audio encoder (processes mel spectrograms)
|
| 42 |
-
- `
|
| 43 |
-
- `decoder_with_past_model_quantized.onnx` - Optimized decoder with KV caching
|
| 44 |
|
| 45 |
### Configuration Files
|
| 46 |
- `config.json` - Model configuration
|
|
@@ -48,41 +40,6 @@ Each zip file contains 7 files needed for inference:
|
|
| 48 |
- `preprocessor_config.json` - Audio preprocessing settings
|
| 49 |
- `tokenizer.json` - Tokenizer vocabulary
|
| 50 |
|
| 51 |
-
## Usage
|
| 52 |
-
|
| 53 |
-
### C# with ONNX Runtime
|
| 54 |
-
```csharp
|
| 55 |
-
// Download and extract zip
|
| 56 |
-
var modelPath = "path/to/extracted/model/";
|
| 57 |
-
|
| 58 |
-
// Initialize with DirectML support
|
| 59 |
-
var sessionOptions = new SessionOptions();
|
| 60 |
-
sessionOptions.AppendExecutionProvider_DML(0);
|
| 61 |
-
|
| 62 |
-
var encoderSession = new InferenceSession(
|
| 63 |
-
Path.Combine(modelPath, "encoder_model_quantized.onnx"), sessionOptions);
|
| 64 |
-
var decoderSession = new InferenceSession(
|
| 65 |
-
Path.Combine(modelPath, "decoder_with_past_model_quantized.onnx"), sessionOptions);
|
| 66 |
-
```
|
| 67 |
-
|
| 68 |
-
### Python with ONNX Runtime
|
| 69 |
-
```python
|
| 70 |
-
import onnxruntime as ort
|
| 71 |
-
|
| 72 |
-
# Load with DirectML/CUDA support
|
| 73 |
-
providers = ['DmlExecutionProvider', 'CPUExecutionProvider']
|
| 74 |
-
encoder_session = ort.InferenceSession('encoder_model_quantized.onnx', providers=providers)
|
| 75 |
-
decoder_session = ort.InferenceSession('decoder_with_past_model_quantized.onnx', providers=providers)
|
| 76 |
-
```
|
| 77 |
-
|
| 78 |
-
## Features
|
| 79 |
-
|
| 80 |
-
✅ **DirectML Support** - Works with any DirectX 12 GPU (AMD, Intel, NVIDIA)
|
| 81 |
-
✅ **CUDA Support** - Accelerated inference on NVIDIA GPUs
|
| 82 |
-
✅ **CPU Fallback** - Automatic fallback to CPU if GPU unavailable
|
| 83 |
-
✅ **Quantized** - INT8/INT4 quantization for smaller size and faster inference
|
| 84 |
-
✅ **Complete** - All files needed for inference included
|
| 85 |
-
|
| 86 |
## Model Sources
|
| 87 |
|
| 88 |
These models are repackaged from:
|
|
|
|
| 19 |
|
| 20 |
| Model | Language | Size | Target Use | Download |
|
| 21 |
|-------|----------|------|------------|----------|
|
| 22 |
+
| **Small English** | English-only | 107MB | Fast English transcription | [whisper-small-en-onnx.zip](small-en/whisper-small-en-onnx.zip) |
|
| 23 |
+
| **Small Multilingual** | 99 languages | 245MB | Fast multilingual transcription | [whisper-small-multilingual-onnx.zip](small-multilingual/whisper-small-multilingual-onnx.zip) |
|
| 24 |
+
| **Medium English** | English-only | 247MB | High quality English transcription | [whisper-medium-en-onnx.zip](medium-en/whisper-medium-en-onnx.zip) |
|
| 25 |
+
| **Medium Multilingual** | 99 languages | 602MB | High quality multilingual | [whisper-medium-multilingual-onnx.zip](medium-multilingual/whisper-medium-multilingual-onnx.zip) |
|
| 26 |
+
| **Large v3 Turbo** | 99 languages | 646MB | Best quality, fastest large model | [whisper-large-v3-turbo-onnx.zip](large-v3-turbo/whisper-large-v3-turbo-onnx.zip) |
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
## Contents of Each Zip
|
| 30 |
|
| 31 |
+
Each zip file contains 6 files needed for inference:
|
| 32 |
|
| 33 |
### ONNX Model Files
|
| 34 |
- `encoder_model_quantized.onnx` - Audio encoder (processes mel spectrograms)
|
| 35 |
+
- `decoder_with_past_model_quantized.onnx` - Text decoder (generates transcription), optimized decoder with KV caching
|
|
|
|
| 36 |
|
| 37 |
### Configuration Files
|
| 38 |
- `config.json` - Model configuration
|
|
|
|
| 40 |
- `preprocessor_config.json` - Audio preprocessing settings
|
| 41 |
- `tokenizer.json` - Tokenizer vocabulary
|
| 42 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
## Model Sources
|
| 44 |
|
| 45 |
These models are repackaged from:
|