jb2n95/model-mirror-tonic

Supertonic-2 TTS Model Mirror

Mirror of Supertone/supertonic-2 with additional float16 quantized version.

Model Versions

Version	Path	Total Size	Description
fp32 (Original)	fp32/	~251MB	Original float32 precision
fp16 (Quantized)	fp16/	~126MB	Float16 quantized (~50% smaller)

Directory Structure

├── fp32/                    # Original float32 version
│   ├── onnx/
│   │   ├── duration_predictor.onnx (1.5MB)
│   │   ├── text_encoder.onnx (26MB)
│   │   ├── vector_estimator.onnx (126MB)
│   │   ├── vocoder.onnx (97MB)
│   │   ├── tts.json
│   │   └── unicode_indexer.json
│   └── voice_styles/
│       ├── F1.json ~ F5.json (Female voices)
│       └── M1.json ~ M5.json (Male voices)
│
└── fp16/                    # Float16 quantized version
    ├── onnx/
    │   ├── duration_predictor.onnx (0.8MB)
    │   ├── text_encoder.onnx (13MB)
    │   ├── vector_estimator.onnx (63MB)
    │   ├── vocoder.onnx (48MB)
    │   ├── tts.json
    │   └── unicode_indexer.json
    └── voice_styles/
        └── (same as fp32)

Usage

Clone specific version

# Clone entire repository
git clone https://huggingface.co/jb2n95/model-mirror-tonic

# Or use git sparse-checkout for specific version
git clone --filter=blob:none --sparse https://huggingface.co/jb2n95/model-mirror-tonic
cd model-mirror-tonic
git sparse-checkout set fp16  # or fp32

In Docker

# For fp32 (original)
RUN git clone --depth 1 https://huggingface.co/jb2n95/model-mirror-tonic && \
    mv model-mirror-tonic/fp32/* public/assets/ && \
    rm -rf model-mirror-tonic

# For fp16 (smaller, faster download)
RUN git clone --depth 1 https://huggingface.co/jb2n95/model-mirror-tonic && \
    mv model-mirror-tonic/fp16/* public/assets/ && \
    rm -rf model-mirror-tonic

Changes from Original

• fp16 version: Converted from float32 to float16 using onnxconverter-common
• No other modifications to model architecture or weights

License

This model is distributed under the OpenRAIL-M License, same as the original.

• Original model: Supertone/supertonic-2
• License: OpenRAIL-M

Attribution

• Original model by Supertone Inc.
• Float16 quantization by jb2n95

Supported Languages

• Korean (ko)
• English (en)
• Spanish (es)
• Portuguese (pt)
• French (fr)