DESIGN AND IMPLEMENTATION OF AI-POWERED SPEECH TO TEXT SYSTEM USING DEEP NEURAL NETWORK

Speech-to-text (STT) systems have gained significant traction in recent years due to their ability to transcribe spoken language into text format, facilitating communication and accessibility for individuals with hearing impairments or those who prefer hands-free interaction. This project addresses the challenges by proposing a modular deep neural network based architecture for robust speech to text transcription, the system design incorporates multiple real-world speech augmentation techniques like background noises, room impulse responses and speed perturbations during training. Novel speaker adaptation methods are introduced to improve generalization across different users without needing explicit enrollment data. The system is implemented using Pytorch and Google Cloud Speech APIs. Rigorous testing on internal datasets as well as public domain LibriSpeech utterances reveal significant improvements compared to existing solutions in the literature. Word error rates drop absolutely across noisy and accented speech without loss in responsiveness. • The modular architecture developed allows seamless deployment across application domains through configuration changes rather than full retraining. The outcomes demonstrate the feasibility of real-time, accurate and personalized speech recognition services even with limited domain data.