VianLB/coqui_tts_fine-tuned

XTTS v2 Fine-tuned Voice Model

This is a fine-tuned XTTS v2 model trained for voice cloning.

Model Details

• Base Model: XTTS v2 (Coqui TTS)
• Training: 400 epochs
• Loss Improvement: 57% reduction (2.863 → 1.231)
• Languages: English & French
• Sample Rate: 24kHz

Files

• best_model.pth - Fine-tuned GPT model weights
• config.json - Model configuration
• vocab.json - Tokenizer vocabulary
• dvae.pth - Discrete VAE for audio encoding
• mel_stats.pth - Mel-spectrogram normalization stats
• reference.wav - Voice reference sample

Usage

from TTS.api import TTS

# Load the model
tts = TTS(model_path="path/to/model/folder", config_path="path/to/config.json")

# Generate speech
tts.tts_to_file(
    text="Your text here",
    file_path="output.wav",
    speaker_wav="reference.wav",
    language="en"
)

Requirements

TTS>=0.22.0
torch>=2.0.0
torchaudio

Training Details

• Dataset: Custom voice recordings
• Training Duration: ~50 hours total
• Batch Size: 2
• Gradient Accumulation: 126 steps
• Optimizer: AdamW

Limitations

• Optimized for the specific voice in the training data
• Best performance on texts similar to training distribution
• May have reduced quality on very long texts

License

Apache 2.0