SNAP Piper Korean Multi-Speaker TTS (22.05kHz, 22 Speakers)

Multi-scale Mel-spectrogram loss fine-tuned high-fidelity VITS ONNX models for Korean speech synthesis. Trained on AI Hub Korean reading dataset with SNAP Frontend Engine phonology & text normalization.

Model Highlights

• Architecture: VITS (Piper) + Multi-Scale Mel Loss Fine-tuning
• Sampling Rate: 22,050 Hz
• Speakers: 22 Voice Actors (11 Male, 11 Female) across age groups 20s to 60s
• Performance: High-speed Real-time inference on CPU with ONNXRuntime (~500+ FPS)
• Checkpoints:
  • Male: Step 9,000 Sweet-spot (HNR top 1, Mel L1 3.289, Whisper WER 0%)
  • Female: Step 8,000 Sweet-spot (HNR 2.09 dB, DynRange 101.9 dB, Whisper WER 0%)

Files

• snap_male11_reading_22k.onnx & .json : Male 11-speaker FP32 ONNX
• snap_female11_reading_22k.onnx & .json : Female 11-speaker FP32 ONNX
• snap_male11_reading_22k_fp16.onnx : Male 11-speaker FP16 ONNX (~58MB)
• snap_female11_reading_22k_fp16.onnx : Female 11-speaker FP16 ONNX (~58MB)

Demo

Experience live synthesis on Hugging Face Spaces: softguy777/snap_voice_demo2