Low latency modeling of temporal contexts for speech recognition

This thesis focuses on the development of neural network acoustic models for large vocabulary continuous speech recognition (LVCSR) to satisfy the design goals of low latency and low computational complexity. Low latency enables online speech recognition; and low computational complexity helps reduce the computational cost both during training and inference. Long span sequential dependencies and sequential distortions in the input vector sequence are a major challenge in acoustic modeling. Recurrent neural networks have been shown to effectively model these dependencies. Specifically, bidirectional long short term memory (BLSTM) networks, provide state-of-the-art performance across several LVCSR tasks. However the deployment of bidirectional models for online LVCSR is non-trivial due to their large latency; and unidirectional LSTM models are typically preferred. In this thesis we explore the use of hierarchical temporal convolution to model long span temporal dependencies. We propose a sub-sampled variant of these temporal convolution neural networks, termed time-delay neural networks (TDNNs). These sub-sampled TDNNs reduce the computation complexity by ~5x, compared to TDNNs, during frame randomized pre-training. These models are shown to be effective in modeling long-span temporal contexts, however there is a performance gap compared to (B)LSTMs. As recent advancements in acoustic model training have eliminated the need for frame randomized pre-training we modify the TDNN architecture to use higher sampling rates, as the increased computation can be amortized over the sequence. These variants of sub- sampled TDNNs provide performance superior to unidirectional LSTM networks, while also affording a lower real time factor (RTF) during inference. However we show that the BLSTM models outperform both the TDNN and LSTM models. We propose a hybrid architecture interleaving temporal convolution and LSTM layers which is shown to outperform the BLSTM models. Further we improve these BLSTM models by using higher frame rates at lower layers and show that the proposed TDNN- LSTM model performs similar to these superior BLSTM models, while reducing the overall latency to 200 ms. Finally we describe an online system for reverberation robust ASR, using the above described models in conjunction with other data augmentation techniques like reverberation simulation, which simulates far-field environments, and volume perturbation, which helps tackle volume variation even without gain normalization