Audio2textNet: A Deep Learning Framework for Robust Audio-to-Text Transcription Using CNNs and Transformers

Audio-to-text transcriptions are becoming increasingly popular as media and other materials continue to be produced and used as more conventional techniques are found to have shortcomings in terms of noise, variability of speakers, and accents. Real-time transcription of the audio back to text, especially when different speakers and backgrounds, presents challenges to this date. In this work, the authors present an original deep learning architecture called Audio2textNet that employs CNN for accurate feature extraction, and transformer for context awareness. The technique involves enhanced data augmentation routines, parameter searches for ideal models, and tough testing on the LibriSpeech corpus. The performance evaluation of Audio2textNet shows that it has better results with 93.8% accuracy, 5.2% WER, and 3.1% CER as compared to other similar systems such as Deep Speech and Wav2Vec. These findings impressively demonstrate the noise insensitivity, speaking variation tolerance, and computational effectiveness in the realtime mode of the suggested model. The proposed system has important implications for accessibility tools, automatic captioning, and multilingual transcription systems thus offering a revolutionary approach to solving existing problems in audio-totext translation.