Hindi-ASR-using-Wave2Vec2

Hindi Speech Recognition with Wav2Vec2

This project fine-tunes the facebook/wav2vec2-large-xlsr-53 model for Hindi speech recognition using the Common Voice dataset.

Overview & Applications

Automatic Speech Recognition (ASR) enables machines to understand and transcribe spoken language into text. This project trains a deep learning model to convert Hindi speech into text, making it useful for:

• Voice Assistants: Enhancing AI-driven virtual assistants with Hindi language support.
• Transcription Services: Automating Hindi audio-to-text conversion for interviews, lectures, and media content.
• Accessibility Tools: Assisting users with hearing impairments by providing real-time speech transcription.
• Language Learning: Helping learners improve pronunciation and comprehension through speech-to-text feedback.
• Voice-Controlled Applications: Integrating with smart devices to process spoken Hindi commands.

Features

• Data Preprocessing: Cleans and prepares audio-text pairs, removing special characters and normalizing text.
• Tokenization & Feature Extraction: Uses a custom vocabulary and Wav2Vec2Processor to convert speech into model-ready input.
• Model Fine-tuning: Trains the pre-trained Wav2Vec2 model on Hindi speech data to improve speech-to-text accuracy.
• Saving & Loading Model: Saves trained model weights for reuse in inference or further training.

Usage

Train the model in Google Colab and download the weights locally after training. The trained model can be used for inference on new audio samples.

Model Checkpoint

After training, the model weights are saved and can be loaded for further fine-tuning or inference.

Acknowledgments

• Mozilla Common Voice Dataset
• Hugging Face Transformers Library