Arabic Audio-to-Text Pipeline with Speaker Diarization

A comprehensive Python pipeline that processes Arabic audio content from YouTube videos, performing speaker diarization, transcription, and text summarization.

Features

• YouTube audio extraction and processing
• Audio denoising for improved quality
• Speaker diarization (identifies different speakers)
• Arabic speech transcription using Faster Whisper
• Text summarization with timestamps
• Multiple output formats (TXT & DOCX)

Prerequisites

Required Software

• Python 3.8 or higher
• FFmpeg (for audio processing)
• CUDA-capable GPU (optional, for faster processing)

Required Accounts

• HuggingFace Account (for model access)
  • Need to accept terms for pyannote/speaker-diarization-3.1
  • Generate and save your HuggingFace token

Installation

1. Clone the repository:

git clone (https://github.com/SultanKhalid92/Speech2text_arabic.git)
cd Speech2text_arabic

2. Install dependencies:

pip install -r requirements.txt

3. Configure settings in settings.yaml:

youtube_url: "YOUR_YOUTUBE_URL"
audio_path: "audio.wav"
clean_audio_path: "audio_denoised.wav"
transcript_path: "transcript_whisper_large.txt"
summary_path: "summary"
hf_token: "YOUR_HUGGINGFACE_TOKEN"
ffmpeg_location: "PATH_TO_FFMPEG"

🚀 Usage

Using Python Script

Run the main script:

python main.py

Using Jupyter Notebook

Open and run Speech2text.ipynb sequentially.

🔧 Pipeline Components

1. Audio Extraction

• Uses yt_dlp to download YouTube audio
• Converts to WAV format at 16kHz
• Handled by extract_audio() function

2. Audio Denoising

• Uses librosa and noisereduce
• Improves audio quality for better transcription
• Implemented in denoise_audio() function

3. Speaker Diarization

• Uses pyannote.audio model
• Identifies different speakers in the audio
• Provides speaker timestamps
• Implemented in run_diarization() function

4. Transcription

• Uses faster-whisper model
• Optimized for Arabic language
• Includes timestamp segmentation
• Handled by transcribe_audio() function

5. Text Processing

• Combines speaker information with transcription
• Generates structured output with timestamps
• Creates both detailed and summarized versions

Output Files