Speech Recognition Project

Overview

This project implements a speech recognition system using Python and the Wav2Vec2 model from Hugging Face. The system is fine-tuned on the PolyAI MINDS-14 dataset and deployed with a real-time transcription interface using Gradio.

Features

• Dataset: Utilizes the PolyAI MINDS-14 dataset, resampled to 16kHz.
• Model & Processor: Employs the Wav2Vec2ForCTC model and processor for audio-to-text transcription.
• Training: Fine-tunes the model with custom training arguments and a data collator for handling varying input lengths.
• Evaluation: Assesses model performance using the Word Error Rate (WER) metric.
• Deployment: Provides a Gradio interface for real-time transcription.

Skills Used

• Python
• Machine Learning
• Deep Learning
• Natural Language Processing (NLP)
• Audio Processing
• PyTorch
• Transformers (Hugging Face)
• Gradio Interface Development
• Model Evaluation (WER)
• Dataset Management
• GPU Acceleration (CUDA)

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/speech-recognition-project.git

2. Navigate to the project directory:

   cd speech-recognition-project

3. Install the required packages:

   pip install -r requirements.txt

Usage

Training the Model

1. Run the training script to fine-tune the model:

   python train.py

Real-Time Transcription

1. Launch the Gradio interface:

   python app.py

2. Use the microphone input to test real-time transcription.

Example

Transcription for a sample audio:

The input test audio is: "Sample transcription"
The output prediction is: "Sample prediction"

Acknowledgements

• Hugging Face for providing the Wav2Vec2 model and transformers library.
• PolyAI for the MINDS-14 dataset.
• Gradio for the interactive interface.