Simple TTS Model

Model Description

• Transformer encoder for text encoding
• Duration predictor
• Transformer decoder for mel spectrogram generation

Training

• Dataset: LJSpeech (5000 samples)
• Hardware: Kaggle T4 GPU
• Training time: 20 epochs

Model Parameters

• Total parameters: 5,168,465
• Hidden dimension: 256
• Number of layers: 3
• Attention heads: 4

Usage

Limitations

• Training on more data
• Adding a vocoder (e.g., HiFi-GAN) for audio generation
• Using phoneme-based input instead of characters