Arabic Automatic Speech Recognition (ASR) System from Scratch

Developed an end-to-end Arabic Automatic Speech Recognition (ASR) system from scratch using PyTorch, without relying on pretrained models.

The system uses a custom CNN + BiLSTM + CTC architecture and was trained on the Mozilla Common Voice Arabic dataset. I implemented the complete pipeline, including audio preprocessing, Arabic text normalization, SpecAugment, mixed-precision training (AMP), checkpoint management, and real-time WER/CER evaluation.

A Greedy CTC Decoder was implemented for speech-to-text inference, along with an interactive Gradio interface for testing live Arabic speech transcription.

Model Performance:

• WER: 35.48%

• CER: 10.69%

Tech Stack:

• PyTorch

• torchaudio

• librosa

• Gradio

• jiwer

This project strengthened my practical experience in deep learning, speech processing, sequence modeling, CTC-based training, and building complete AI systems from the ground up.