RAttSR: A Novel Low-Cost Reconstructed Attention-Based End-to-End Speech Recognizer

People are curious about voice commands for the next generation of interaction. It will play a dominant role in communicating with smart devices in the future. However, language remains a significant barrier to the widespread use of these devices. Even the existing models for the traditional languages need to compute extensive parameters, resulting in higher computational costs. The most inconvenient in the latest advanced models is that they are unable to function on devices with constrained resources. This paper proposes a novel end-to-end speech recognition based on a low-cost Bidirectional Long Short Term Memory (BiLSTM) attention model. The mel-spectrogram of the speech signals has been generated to feed into the proposed neural attention model to classify isolated words. It consists of three convolution layers followed by two layers of BiLSTM that encode a vector of length 64 to get attention against the input sequence. The convolution layers characterize the relationship among the energy bins in the spectrogram. The BiLSTM network removes the prolonged reliance on the input sequence, and the attention block finds the most significant region in the input sequence, reducing the computational cost in the classification process. The encoded vector by the attention head is fed to three-layered fully connected networks for recognition. The model takes only 133K parameters, less than several current state-of-the-art models for isolated word recognition. Two datasets, the Speech Command Dataset (SCD), and a self-made dataset we developed for fifteen spoken colors in the Bengali dialect, are utilized in this study. Applying the proposed technique, the performance evaluation with validation and test accuracy in the Bengali color dataset reaches 98.82% and 98.95%, respectively, which outperforms the current state-of-the-art models regarding accuracy and model size. When the SCD has been trained using the same network model, the average test accuracy obtained is 96.95%. To underpin the proposed model, the outcome is compared with the recent state-of-the-art models, and the result shows the superiority of the proposed model.