MiraTTS-onnx-int4

MiraTTS ONNX INT4

This is an INT4 quantized ONNX version of YatharthS/MiraTTS, optimized for efficient CPU and GPU inference using ONNX Runtime with the GenAI extension.

Key Benefits

• ~2x faster inference compared to FP32 ONNX
• ~15x less memory usage (857MB vs 2.5GB+ for original)
• CPU-friendly: Runs efficiently on CPU without GPU
• Same quality: Negligible quality loss from quantization

Model Details

Usage

Installation

pip install onnxruntime-genai torch soundfile

For GPU support:

pip install onnxruntime-genai-cuda torch soundfile

Python API

from huggingface_hub import snapshot_download
from mira.model import MiraTTS
import soundfile as sf

# Download model
model_path = snapshot_download("uetuluk2/MiraTTS-onnx-int4")

# Load model (use device="cuda" for GPU)
tts = MiraTTS(model_path, device="cpu")

# Encode reference audio for voice cloning
context = tts.encode_audio("reference.wav")

# Generate speech
audio = tts.generate("Hello, this is a test!", context)

# Save output
sf.write("output.wav", audio.cpu().numpy(), 48000)

With the MiraTTS-onnx Repository

git clone https://github.com/uetuluk/MiraTTS-onnx
cd MiraTTS-onnx
pip install -e .

# Download model
huggingface-cli download uetuluk2/MiraTTS-onnx-int4 --local-dir ./models/int4

# Run example
python examples/generate_int4.py --text "Hello world!" --reference your_voice.wav

Performance

Tested on various hardware configurations:

Quantization

This model was quantized using onnxruntime-genai with INT4 weight quantization:

python -m onnxruntime_genai.models.builder \
    -m YatharthS/MiraTTS \
    -e cpu \
    -p int4 \
    -o ./mira_onnx_int4

License

This model inherits the CC BY-NC-SA 4.0 license from the original MiraTTS model.

Credits

• Original model by YatharthS
• INT4 quantization by uetuluk2
• ONNX conversion code: MiraTTS-onnx