Hands-On Deep Learning for IoT

DL models for ASR

A number of DL algorithms or models have been used in ASR. A Deep Belief Network (DBN) is one of the early implementations of DL in ASR. Generally, it has been used as a pre-training layer with a single supervised layer of a Deep Neural Network (DNN). Long Short-Term Memory (LSTM) has been used for large-scale acoustic modeling. Time Delay Neural Network (TDNN) architectures have been used for audio signal processing. CNN, which has popularized DL, is also used as DL architecture for ASR. Use of DL architectures has significantly improved the speech recognition accuracy of ASRs. However, not all DL architectures have shown improvements, especially in different types of audio signals and environments, such as noisy and reverberant ...