Stephen Jie Hoon Koo - Academia.edu (original) (raw)
Uploads
Papers by Stephen Jie Hoon Koo
Automatic speech recognition (ASR) has been a subject of research by linguists and computer scien... more Automatic speech recognition (ASR) has been a subject of research by linguists and computer scientists for decades. In recent years, state-of-theart speech recognition systems have moved from HMM-based systems to end-to-end neural systems. These systems usually run on powerful machines or clusters. However one of the biggest usages of automatic speech recognition is in applications running on mobile and wearable devices, requiring acoustic data to be sent from the device to data centers. In many situations, it is desirable to have a speech recognition system that performs relatively well offline as well. Offline ASR systems have the benefits of reduced latency as well as improved reliability, but face the challenges of a memory, compute, and power-constrained environment. In our project, we focus on investigating model compression methods for reducing the memory and computational footprint of deep speech recognition networks.
Automatic speech recognition (ASR) has been a subject of research by linguists and computer scien... more Automatic speech recognition (ASR) has been a subject of research by linguists and computer scientists for decades. In recent years, state-of-theart speech recognition systems have moved from HMM-based systems to end-to-end neural systems. These systems usually run on powerful machines or clusters. However one of the biggest usages of automatic speech recognition is in applications running on mobile and wearable devices, requiring acoustic data to be sent from the device to data centers. In many situations, it is desirable to have a speech recognition system that performs relatively well offline as well. Offline ASR systems have the benefits of reduced latency as well as improved reliability, but face the challenges of a memory, compute, and power-constrained environment. In our project, we focus on investigating model compression methods for reducing the memory and computational footprint of deep speech recognition networks.