VoxCPM Spanish ONNX Models (18k Steps)
This repository contains ONNX exported models for the VoxCPM Spanish fine-tune (18k steps).
Contents
- fp32_v2/: Full precision FP32 models (Best quality)
- int8_v2/: Quantized INT8 models (Fastest, suitable for mobile/CPU)
- scripts/: Inference scripts for both versions
Usage
FP32 Inference
python scripts/test_onnx_inference_v2.py --onnx-dir fp32_v2 --text "Hola mundo"
INT8 Inference
python scripts/test_onnx_inference_int8.py --onnx-dir int8_v2 --text "Hola mundo"
Performance (CPU)
- FP32: ~56x RTF
- INT8: ~52x RTF (Similar quality to FP32, much smaller size)