This repository contains an Automatic Speech Recognition (ASR) system for the Telugu language, leveraging OpenAI's Whisper model and PyTorch. The project involves preprocessing Telugu speech data, fine-tuning the Whisper model, and optimizing it for mobile deployment.
- Preprocessing Telugu Speech Data**: Noise removal and silence trimming
- ASR Model: Using Whisper-small/tiny (PyTorch version)
- Fine-tuning: Training Whisper on Telugu audio dataset
- Model Optimization: Converting Whisper to TorchScript or ONNX for mobile deployment
- Python 3.x
- PyTorch
- OpenAI Whisper
- Hugging Face Transformers
- NumPy, Pandas
- Librosa (for audio processing)
The dataset consists of Telugu audio files and a corresponding TSV text file. The files are stored in https://drive.google.com/drive/folders/1ZFnbfTRbvldqd_igEYvKKuddZDZh2sgJ?usp=sharing in Google Drive.
- Model performance evaluation on Telugu speech data
- Accuracy, WER (Word Error Rate), and loss metrics
- Enhance dataset size and quality
- Implement real-time speech-to-text inference
- Improve inference speed on mobile devices
- Hugging Face for ASR fine-tuning resources