Real Voice Cloning Implementation

New Features:

• WavLM speaker encoder for real voice extraction
• 192-dim to 512-dim embedding projection
• Audio preprocessing (16kHz resample, normalization)
• Voice preview before generation
• Blend ratio for voice stability (70% custom + 30% default)
• Toggle between default voice and cloned voice

How Voice Cloning Works:

1. Upload audio file (5-30 seconds, clear voice)
2. Extract features using WavLM-base-plus-sv model
3. Project 192-dim embeddings to 512-dim SpeechT5 space
4. Normalize and blend with default for stability
5. Use custom embeddings for generation

Technology:

• Speaker Encoder: Xenova/wavlm-base-plus-sv
• Feature pooling: Mean pooling with normalization
• Projection: Linear interpolation with normalization
• Blend: 70% custom + 30% default for quality