EgyptianArabic-Phoneme-VITS

Phoneme-Based TTS for Egyptian Arabic (Training VITS from Coqui)

This repository contains the workflow for training a text-to-speech (TTS) system for Egyptian Arabic using the VITS model from Coqui TTS. It includes dataset preparation, model configuration, and training steps tailored for phoneme-based synthesis.

Features

• Dataset Preparation: Prepares the dataset, including phoneme-based text and corresponding audio files, for training.
• Model Training: Implements the VITS model training pipeline.
• Phoneme-Based Synthesis: Focuses on generating high-quality phoneme-aligned TTS outputs.

Requirements

The notebook was developed using Google Colab. Below are the key dependencies:

• Python 3.x
• Coqui TTS
• ffmpeg
• Additional libraries: zipfile, os, subprocess

Setup Instructions

Dataset Preparation:

1. Mount Google Drive to access the dataset using the provided script:

   from google.colab import drive
   drive.mount("/content/drive")
   

2. Extract the dataset ZIP file and organize the WAV files for processing:

   import os
   import subprocess
   
   input_path = "/content/final_VITS_dataset/wavs"
   output_path = "/content/VITS_processed_dataset/wavs"
   os.makedirs(output_path, exist_ok=True)
   

Model Path:

https://drive.google.com/file/d/1N4ZPviPghoYqvqpNKW2fHrCJ96qNEaDy/view?usp=sharing

Model Inference:

from TTS.utils.synthesizer import Synthesizer

# Updated parameters for the synthesizer
model_path = "/content/drive/MyDrive/VITS-TRIAL-PHON-2/runs/VITS-AR-phon-2-December-23-2024_12+04AM-0000000/checkpoint_58000.pth"
config_path = "/content/drive/MyDrive/VITS-TRIAL-PHON-2/runs/VITS-AR-phon-2-December-23-2024_12+04AM-0000000/config.json"

# Initialize the synthesizer
synthesizer = Synthesizer(
    tts_checkpoint=model_path,
    tts_config_path=config_path,
    vocoder_checkpoint=None,
)

# Adjust speech parameters for slower, clearer output
synthesizer.tts_model.args.length_scale = 2 # Increase for slower speech
synthesizer.tts_model.args.inference_noise_scale = 0.5  # Reduce noise
synthesizer.tts_model.args.inference_noise_scale_dp = 0.8  # Reduce noise distortions

# Text to synthesize
text = "عامل ايه يا حبيبي واحشني أوي فينك من زمان كده"

# Generate speech
wav = synthesizer.tts(text)

# Save the output to a file
output_path = "output_slow_clear.wav"
synthesizer.save_wav(wav, output_path)

print(f"Audio saved to {output_path}")

from IPython.display import Audio
Audio(output_path, autoplay=True)