simon19f3/Voice-Smoother

AI model that detects and remove Amharic filler words from conversation Amharic Voice Smoother A Python-based system to enhance Amharic speech by detecting and removing filler words (e.g., "እሰይ", "እንትን") from audio, improving clarity for digital education and accessibility. Built with speechbrain, transformers, pydub, and torch, it leverages Google Colab’s T4 GPU. Features Silence-based audio segmentation Amharic transcription using speechbrain/asr-wav2vec2-dvoice-amharic Filler detection with fine-tuned Davlan/afro-xlmr-base (75% accuracy, 70% F1-score) Cleaned audio output in WAV format JSON/CSV dataset generation Scalable batch processing (~210s for 30min audio) Requirements Python 3.8+ Dependencies: pip install speechbrain transformers pydub torch pandas tqdm gdown datasets FFmpeg: apt update && apt install -y ffmpeg Hardware: Laptop (2.4 GHz quad-core, 8GB RAM) or Google Colab with T4 GPU Google Drive access for datasets Installation git clone github.com cd amharic-voice-smoother pip install -r requirements.txt apt update && apt install -y ffmpeg Usage Place audio files (WAV, MP3, M4A, OGG) in audio_files/ or use Google Drive links: Training: 1-9SZiAmR3IKuoSdYIkStwRJSoA0mQANR Test: 1-Gv8Ha2ejS_LpX-hFtJuaJ8OmyLYcq-d Run:python main.py Outputs: Cleaned audio: output_clean.wav Datasets: test_filler_word_dataset.json, test_filler_word_dataset.csv Model: amharic_filler_model/ Directory Structure amharic-voice-smoother/ ├── audio_files/ # Input audio ├── pretrained_models/ # ASR model weights ├── amharic_filler_model/ # Trained transformer model ├── test_filler_word_dataset.json # JSON dataset ├── test_filler_word_dataset.csv # CSV dataset ├── output_clean.wav # Cleaned audio ├── main.py # Main script ├── README.md # This file Performance Filler Detection: 75% accuracy, 70% F1-score Processing: 22.6s/min audio (GPU) Clarity: 3.0/5 (native speaker rating) Scalability: 30min batch in ~210s Limitations …