Hybrid deep learning based automatic speech recognition model for recognizing non-Indian languages

Speech is a natural phenomenon and a significant mode of communication used by humans that is divided into two categories, human-to-human and human-to-machine. Human-to-human communication depends on the language the speaker uses. In contrast, human-to-machine communication is a technique in which machines recognize human speech and act accordingly, often termed Automatic Speech Recognition (ASR). Recognition of Non-Indian language is challenging due to pitch variations and other factors such as accent, pronunciation, etc. This paper proposes a novel approach based on Dense Net201 and EfficientNetB0, i.e., a hybrid model for the recognition of Speech. Initially, 76,263 speech samples are taken from 11 non-Indian languages, including Chinese, Dutch, Finnish, French, German, Greek, Hungarian, Japanese, Russian, Spanish and Persian. When collected, these speech samples are pre-processed by removing noise. Then, Spectrogram, Short-Term Fourier Transform (STFT), Spectral Rolloff-Bandwidth, Mel-frequency Cepstral Coefficient (MFCC), and Chroma feature are used to extract features from the speech sample. Further, a comparative analysis of the proposed approach is shown with other Deep Learning (DL) models like ResNet10, Inception V3, VGG16, DenseNet201, and EfficientNetB0. Standard parameters like Precision, Recall, F1-Score, Confusion Matrix, Accuracy, and Loss curves are used to evaluate the performance of each model by considering speech samples from all the languages mentioned above. Thus, the experimental results show that the hybrid model stands out from all the other models by giving the highest recognition accuracy of 99.84% with a loss of 0.004%.