Arabic Sales TTS - OuteTTS Finetuning

This project fine-tunes the OuteTTS v1.0 model specifically for Arabic sales speech synthesis using Unsloth optimization.

Project Overview

OuteTTS is a text-to-speech model that can generate high-quality natural speech. This project adapts OuteTTS v1.0 specifically for Arabic sales conversations by fine-tuning on domain-specific data.

Key features:

• Fine-tuning OuteTTS v1.0 (1B parameter model) on Arabic data
• Using Unsloth optimization for efficient training with lower VRAM requirements
• Processing audio with Whisper for proper word-level alignment
• Supporting speaker identification for multi-speaker datasets

Directory Structure

arabic_tts_finetuning/
├── configs/             # Configuration files
├── data/                # Data utilities and preprocessing
├── models/              # Model training and inference
├── src/                 # Core source code
│   ├── preprocessing/   # Audio and text preprocessing utilities
│   ├── training/        # Training scripts
│   └── inference/       # Inference scripts
└── requirements.txt     # Dependencies

Installation

1. Clone the repository:

git clone https://github.com/yourusername/arabic-tts-finetuning.git
cd arabic-tts-finetuning

2. Install the dependencies:

pip install -r requirements.txt

3. Make sure you have the following packages installed:

• torch
• unsloth
• transformers
• datasets
• whisper
• dac
• outetts
• polars

Usage

Data Preparation

1. Place your audio data in a structure with:

   • WAV files in a wavs directory
   • A metadata CSV file with columns: audio_file|text|speaker_name

2. Run data preprocessing:

python src/preprocessing/process_dataset.py --data_dir /path/to/data --output_dir /path/to/output

Training

python src/training/train.py --config configs/train_config.yaml

Inference

python src/inference/generate.py --model_path /path/to/finetuned/model --text "يوجد شقة للبيع في القاهرة الجديدة بمساحة 120 متر مربع."

Model Architecture

This project uses OuteTTS v1.0 (1B parameters) with the following enhancements:

• Whisper for audio transcription and word alignment
• DAC (Diffusion Audio Codec) instead of WavTokenizer used in older versions
• Interface V3 for modern voice generation
• Unsloth optimization for efficient training

License

This project is licensed under the Apache 2.0 License.