Neural - Arabic ASR
CompletedDeep Learning Arabic Speech Recognition System
Built for a Deep Learning university course. Custom CNN+LSTM model trained from scratch on Mozilla Common Voice Arabic (CTC loss, 30-class Arabic vocab). Integrated with OpenAI Whisper (API + local) and Wav2Vec2-XLSR-Arabic. Full feature set: transcription, speaker diarization, emotion detection (happy/angry/sad/neutral), Arabic text summarization, and semantic voice note search (keyword, single-word, embedding-based). Next.js 16 frontend + FastAPI backend deployed on Railway.
Tech Stack
PythonPyTorchFastAPINext.js 16HuggingFace TransformersOpenAI WhisperWav2Vec2DockerRailway
Key Features
Custom CNN+LSTM Arabic ASR Model (trained from scratch)
Whisper API + Local & Wav2Vec2-XLSR-Arabic
Speaker Diarization
Emotion Detection from Voice
Arabic Text Summarization
Semantic Voice Note Search
Next.js 16 Frontend + FastAPI Backend
Deployed on Railway with Docker
Category: AI