Acoustic-Phonetic Analysis Using the TIMIT Corpus with Traditional and Deep Learning Methods

Abstract

In the paper, a powerful automatic speech recognition system as per the TIMIT corpus is presented, which utilizes a hybrid deep learning model that incorporates convolutional neural networks (CNN), attention through transformers, and long short-term memory layers (BiLSTM). This combined method is made to solve noise-resistant speech recognition: it effectively captures spectral and temporal correlations of speech signals. The accuracy is groundbreaking 99.6% that the framework earns a new level of standards in the sphere. Also, the system is highly noise resilient, with an accuracy of 96% at -5 dB signal-noise ratio (SNR), which is much greater than the former models. One of the most interesting features of this work is that it has a designed pipeline of feature extraction, which includes Mel-frequency cepstral coefficients (MFCCs), as well as delta and delta-delta features. This combination helps to increase the capability of the model to represent both static and dynamic elements of speech. It shows that the architectural advancements are the primary cause of the performance improvements, and the given changes in the conventional framing parameters are not related to its performance. Another strength is that the model can be very accurate in conditions of noise where it can reach up to 89.7% accuracy at -5 dB SNR in comparison to the base line of 73.7%. The study adds to the existing literature on deep learning-based speech recognition as it proposes a model that can be used in a wide range of acoustic conditions without any explicit noise reduction preprocessing’s. The implications of the findings to real-world applications, such as voice-controlled assistants in noisy settings, are substantial, and the findings are a solid base in the further development of the technologies of noisy speech recognition. Future studies will be aimed at performance improvement in phonetically difficult courses and increasing computational efficiency.