Whisper ONNX Converter

Convert OpenAI's Whisper speech recognition models to ONNX format for deployment on Magic Leap 2 and other platforms.

Features

• Convert Whisper models to ONNX format
• Support for all Whisper model sizes
• Character map generation for token decoding
• Optimized for real-time inference
• Unity/Magic Leap 2 deployment support

Installation

git clone git@github.com:miladnasiri/Whisper-ONNX-conver.git
cd Whisper-ONNX-conver
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Usage

Model Conversion

from src.converter import WhisperONNXConverter
converter = WhisperONNXConverter("tiny")
converter.load_model()
converter.convert_to_onnx("whisper_tiny.onnx")

Character Map Generation

from src.tokenizer_utils import WhisperTokenizer
tokenizer = WhisperTokenizer("tiny")
tokenizer.save_char_map("whisper_char_map.json")

Verification

python test_existing_model.py
python check_char_map.py

Model Details

• Input shape: [batch_size, 80, n_frames]
• Output shape: [1, 1500, 384]
• Character map size: 50,363 tokens

Documentation

• Character Map Documentation

Contact

Milad Nasiri