A cloud-based framework for Thai Large Vocabulary Speech Recognition

This paper presents an improvement of a distributed Thai speech recognizer (SR). Two main objectives of the improvement are investigated; 1) the response time in terms of a real-time factor (RTF), 2) the cloud computing deployment. The proposed framework adapts and migrates the baseline collaborative DSR system to the Docker platform. Multiple containers are shared system resources such as CPU, memory, storage and I/O. We separate the services from the baseline framework into Docker containers; a front-end service container and SR engine containers. Message Queue service is initiated by an event-based trigger for controlling and communicating among containers. We also modify the recognition workflow to improve the overall run-time operation by using multiple SR engines to perform the parallel recognition processes by mean of utterance decoding. To support the parallel decoding process, an endpoint detection and voice activity detection are implemented in the client side for defining a chunk of a speech utterance. Meanwhile, a SR engine load balancer is also developed to find an available SR engine container, and assign to the queue respectively. The proposed system is compared with a baseline system regarding their computational efficiency and cost of resources. The results show that the implemented cloud-based framework with the parallel of multiple SR engines can substantially reduce the computational cost concerning the RTF and can handle the higher number of concurrent input on average. Moreover, in the same hardware setting, our proposed framework can support more SR engine than the baseline one. The system infrastructure can be flexible, scalable, and reliable with no significant difference in WER between the proposed and baseline systems.